A public compendium · updated daily
How frontier intelligence is compressed, priced, and contested.
Distillation moves capability from a large teacher model into a small, cheap student. It is the quiet engine behind most models people actually pay for, the subject of an open dispute between the largest labs, and a live regulatory question in three jurisdictions. This compendium tracks all of it.
Customer · compiled 4 September 2026 · 78 sources
The buyer’s view of AI distillation: which small model to actually deploy
For an enterprise buyer in September 2026, the distillation question is no longer "is the small model good enough" but "which small model, and what does the licence let me do with it". Within a single generation, distilled and small-sibling tiers retain 88–98% of their larger sibling's benchmark score for 3–25% of the price: GPT-5.6 Luna holds about 94% of Sol's GPQA Diamond at 5% of the input rate, DeepSeek-V4-Flash scores 98% of V4-Pro's at a third of the price, and Gemini 2.5 Flash — which Google explicitly documents as a k-sparse logit distillation of 2.5 Pro — holds 96% at 24%; across a large parameter gap, retention falls to 47–76%, as DeepSeek's own R1 students and Google's Gemma E-series show. The gap that remains is not knowledge but agentic reliability: on SWE-bench Verified, Claude Sonnet 5 retains 89% of Opus 5 and Haiku 4.5 only 76%, so long-horizon coding and tool-use workloads are the last place a frontier tier still pays for itself. Open weights have become the buyer's real leverage — DeepSeek V4 (MIT), Gemma 4 (Apache 2.0), Qwen3.8-27B (Apache 2.0) and Mistral Small 4 (Apache 2.0) all clear 71.2–90.1 GPQA Diamond with no per-token fee and no anti-distillation clause — which is why the Silicon Data enterprise inference index fell from $2.04/MTok on 31 May 2026 to $1.16–1.18 in early August. The three things that should drive the decision, in order, are: whether your traffic needs frontier-grade agentic reliability on more than 15% of calls; whether the vendor's terms let you distil its outputs into your own model (OpenAI and Anthropic say no, DeepSeek's MIT licence says yes); and whether you can absorb the migration cost when a tier is repriced or retired — which, on 2026 evidence, happens roughly every quarter.
Key figures · 8 figures
Enterprise inference price index
1.17 USD / M tokens
-43% since 31 May 2026
Silicon Data index cited by Jefferies; hit a 2026 low of $1.16–$1.18 on 6–8 Aug 2026, down from $2.04 on 31 May and $1.45 in late July.
Cheapest hosted model per 1M tokens (in)
0.035 USD
Amazon Nova Micro
Amazon's cheapest Nova tier. AWS supports it as a distillation student in Bedrock Model Distillation, but does not state that the shipped weights were distilled from Premier or Pro. Text-only, 128K context.
GPQA retained by GPT-5.6 Luna vs Sol
94.2 %
at 5% of the input price (94.2–97.3% across providers)
87.0 vs 92.4 GPQA Diamond; $0.20 vs $4.00 per M input tokens. Both figures are third-party (OpenRouter auto-routing); provider-specific Luna scores on the same page run to 89.9, and DataLearner reports Sol at 93.5 at max thinking, so retention is best read as a 94.2–97.3% range pending the GPT-5.6 system card.
GPQA of DeepSeek-V4-Flash vs V4-Pro
97.8 %
at 33% of the price
88.1 vs 90.1 GPQA Diamond, both MIT-licensed open weights. Same-family comparison: DeepSeek documents V4-Flash as a consolidation of V4 domain experts via on-policy distillation, not as a distillation of V4-Pro.
SWE-bench Verified retained by Claude Haiku 4.5 vs Opus 5
76.4 %
at 20% of the price
73.3 vs 96.0. Agentic coding is where the distilled tiers still lose most.
Cost per 1M requests, GPT-5.6 Sol vs Luna
18 x cheaper
$3,600 → $200
At 400 input + 100 output tokens per request, list price, no caching.
Best open-weight GPQA Diamond in this dataset
90.1 %
DeepSeek-V4-Pro, MIT licence (open-weight range 71.2–90.1)
1.6T-parameter MoE with 49B active; within 4.2 points of Gemini 3.1 Pro.
Bedrock Model Distillation claim
75 % cheaper
and up to 500% faster
AWS states distilled models are up to 500% faster and up to 75% less expensive than the original, with under 2% accuracy loss on RAG-style use cases.
Key findings · 10 findings
The cheap tier is now good enough for roughly 85% of enterprise traffic
Within a single current generation the small sibling retains 88–98% of its teacher's knowledge benchmark. GPT-5.6 Luna scores 87.0 GPQA Diamond against Sol's 92.4; GPT-5.4 mini scores 88.0 against 93.0; Gemini 2.5 Flash scores 82.8 against 2.5 Pro's 86.4; DeepSeek-V4-Flash scores 88.1 against V4-Pro's 90.1. Across a large parameter gap retention falls to 47–76% — DeepSeek's own R1 students range from 91.2% at 70B down to 47.3% at 1.5B, and Gemma 4 31B to E4B retains 69.5%. The practical rule that falls out of the numbers is to route the hardest 5–15% of traffic to a frontier tier and everything else down to a same-generation small tier, not to the smallest model available.
Agentic coding is the one place the frontier tier still earns its price
Knowledge benchmarks compress; long-horizon agent benchmarks do not. Claude Opus 5 scores 96.0 on SWE-bench Verified, Sonnet 5 85.2 (89% retention) and Haiku 4.5 73.3 (76% retention). Llama 4 Scout retains 92% of Maverick’s MMLU-Pro but only 76% of its LiveCodeBench. If your workload is a coding agent that must finish a multi-step task unsupervised, the retention curve is much steeper than the GPQA curve suggests and a cheap tier will show up as retries, not as wrong answers.
Open weights are the buyer’s only real negotiating position
Four open-weight families now clear 71.2–90.1 GPQA Diamond with no per-token fee and no anti-distillation clause: DeepSeek V4-Pro/Flash (MIT, 90.1/88.1), Qwen3.8-27B (Apache 2.0, 89.2), Gemma 4 31B (Apache 2.0, 84.3 GPQA / 85.2 MMLU-Pro) and Mistral Small 4 (Apache 2.0, 71.2 GPQA / 78.0 MMLU-Pro). Having a credible self-host fallback is what makes an API price cut stick — and 2026 has been a year of price cuts.
You may not distil the models you buy — but you may distil the ones you download
Anthropic’s Commercial Terms section D.4 bars customers from accessing the Services "to build a competing product or service, including to train competing AI models". OpenAI’s Services Agreement carries an equivalent restriction. Both apply to enterprise accounts. By contrast DeepSeek R1 and V4 ship under MIT, and the R1 release explicitly encourages distillation; Qwen3.8, Gemma 4, Mistral Small 4 and Ministral 3 are Apache 2.0. If your roadmap includes training a small in-house model on model outputs, the licence question decides your teacher before any benchmark does.
Distillation is now a purchasable service, not just a research technique
Amazon Bedrock Model Distillation takes only your prompts: it generates synthetic teacher responses and fine-tunes the student. Nova Premier, Claude 3.5 Sonnet v2 and Llama 3.3 70B are supported teachers; Amazon Nova Pro and Llama 3.2 1B/3B are supported students. AWS claims up to 500% faster and up to 75% less expensive inference than the original models, with less than 2% accuracy loss for use cases like RAG. For most buyers this is a cheaper path to a task-specific small model than running a distillation pipeline in-house.
Distillation transfers narrow skills far better than broad knowledge
The DeepSeek R1 student series is the cleanest natural experiment available. R1-Distill-Qwen-1.5B keeps 86% of R1’s MATH-500 (83.9 vs 97.3-class teacher performance) but only 47% of its GPQA Diamond (33.8 vs 71.5). At 7B the split is still stark: MATH-500 92.8 but GPQA 49.1. Buyers evaluating a tiny distilled model on a maths or format-following eval will systematically overestimate how it behaves on open-domain knowledge work.
Advertised latency and measured latency have diverged because of adaptive thinking
Vendor tables still say "fastest", but the number a buyer feels now depends on the reasoning effort setting. Artificial Analysis measures GPT-5.6 Luna at 1.70s time-to-first-token at low effort and 19.87s at high; Claude Haiku 4.5 with reasoning on measures 19.92s; Claude Sonnet 5 at max effort measures 177.77s. Gemini 2.5 Flash-Lite in non-reasoning mode measures 0.30s. Any latency SLA written against a reasoning model has to pin the effort level or it is not a specification.
Legacy tiers are the biggest silent line item in most 2026 AI bills
GPT-4o still lists at $2.50/$10 — the same rate as in 2024 — while GPT-5.6 Luna lists at $0.20/$1.20 with a far larger context window. Claude Sonnet 4.6 costs 50% more than the newer, better Sonnet 5. Gemini 3.5 Flash at $1.50/$9.00 costs twice Gemini 3.8 Flash's $0.75/$3.75 — but that Flash rate is promotional through 2026-12-31 and lists at $1.50/$7.50 from 2027-01-01, i.e. the same input rate as 3.5 Flash. Nothing forces a migration, so pinned model IDs from 2024–25 quietly bill at up to 12x the current market rate for the same job.
Price stability is not something you can assume any more
In 2026 alone: OpenAI cut GPT-5.6 Luna 80% and Terra 20% on 30 July, then cut Sol over 20% on 21 August for a three-month window; Anthropic cancelled a scheduled Sonnet 5 increase from $2/$10 to $3/$15 and made the lower price permanent; DeepSeek raised V4 standard rates roughly 3–4.7x on 16 August — and cache-hit input rates by up to 11x — as demand strained capacity. Contract for the workload, not for the price — and keep a second vendor wired up.
Context window is no longer a reason to pay frontier prices
Every GPT-5.6 tier including the $0.20 Luna carries a 1.05M-token window. Claude Sonnet 5 and Opus 5 both carry 1M. Gemini 3.5 Flash-Lite carries 1M at $0.30/MTok. Llama 4 Scout carries 10M with open weights. Two caveats matter for whole-repository or whole-contract passes: Claude Haiku 4.5 is still capped at 200K, and long prompts are repriced above 200K input tokens — Gemini 3.1 Pro and 2.5 Pro double the input rate above that threshold (and GPT-5.4-class models apply 2x input / 1.5x output), so the cheap headline rate is not the rate you pay on a million-token prompt.
Charts · 6 charts
Blended price vs MMLU-Pro
%| Point | Series | Blended price (USD per million tokens, 3:1 input:output) | MMLU-Pro (%) % |
|---|---|---|---|
| Gemini 3.1 Pro (Preview) | Frontier teachers | 4.5 | 92.6 |
| DeepSeek-V4-Pro | Frontier teachers | 1.98 | 87.5 |
| Gemini 3.1 Flash-Lite | Small siblings | 0.563 | 83 |
| Mistral Small 4 | Small siblings | 0.263 | 78 |
| DeepSeek-V4-Flash | Documented distillations | 0.66 | 86.2 |
| Llama 3.3 70B Instruct | Open weights (hosted price) | 1.04 | 68.9 |
Only models that publish MMLU-Pro AND have a sourced token price appear here. The frontier is almost flat between $0.26 and $2.00: Mistral Small 4 buys 78.0 MMLU-Pro for $0.26 blended, DeepSeek-V4-Flash 86.2 for $0.66, DeepSeek-V4-Pro 87.5 for $1.98, and Gemini 3.1 Pro 92.6 for $4.50 — so the last 6 MMLU-Pro points cost roughly 7x. Gemma 4 31B (85.2) and Gemma 4 26B A4B (82.6) sit off this chart because they are self-host-only and therefore have no list token price.
Sources: ai.google.dev · deepmind.google · layerlens.ai · api-docs.deepseek.com · huggingface.co · huggingface.co · mistral.ai · openrouter.ai · together.ai · github.com
Blended price vs GPQA Diamond — the whole shortlist on one plot
%| Point | Series | Blended price (USD per million tokens, 3:1 input:output) | GPQA Diamond (%) % |
|---|---|---|---|
| GPT-5.6 Sol | Frontier teachers | 8 | 92.4 |
| GPT-5.4 | Frontier teachers | 5.63 | 93 |
| o3 | Frontier teachers | 3.5 | 83.3 |
| Gemini 3.1 Pro (Preview) | Frontier teachers | 4.5 | 94.3 |
| Gemini 2.5 Pro | Frontier teachers | 3.44 | 86.4 |
| DeepSeek-V4-Pro | Frontier teachers | 1.98 | 90.1 |
| Amazon Nova Pro | Frontier teachers | 1.4 | 46.9 |
| GPT-5.6 Terra | Small siblings (method undisclosed) | 4.5 | 88.4 |
| GPT-5.6 Luna | Small siblings (method undisclosed) | 0.45 | 87 |
| GPT-5.4 mini | Small siblings (method undisclosed) | 1.69 | 88 |
| GPT-5.4 nano | Small siblings (method undisclosed) | 0.463 | 82.8 |
| GPT-5 mini | Small siblings (method undisclosed) | 0.688 | 80.3 |
| GPT-5 nano | Small siblings (method undisclosed) | 0.138 | 70.9 |
| o4-mini | Small siblings (method undisclosed) | 1.93 | 81.4 |
| Gemini 3.1 Flash-Lite | Small siblings (method undisclosed) | 0.563 | 72.2 |
| Mistral Small 4 | Small siblings (method undisclosed) | 0.263 | 71.2 |
| Amazon Nova Lite | Small siblings (method undisclosed) | 0.105 | 42 |
| Amazon Nova Micro | Small siblings (method undisclosed) | 0.061 | 40 |
| Gemini 2.5 Flash | Documented distillations | 0.85 | 82.8 |
| DeepSeek-V4-Flash | Documented distillations | 0.66 | 88.1 |
| gpt-oss-20b | Open weights | 0.131 | 58.6 |
| Llama 3.3 70B Instruct | Open weights | 1.04 | 50.5 |
| Qwen3.8-27B | Open weights | 1.13 | 89.2 |
| Ministral 3 14B Instruct | Open weights | 0.2 | 71.2 |
The upper-left corner is the whole story: GPT-5.6 Luna at $0.45 blended / 87.0 GPQA and DeepSeek-V4-Flash at $0.66 / 88.1 sit close to the frontier for a fraction of the price. Gemini 3.1 Flash-Lite is cheaper still at $0.56 blended but scores 72.2 GPQA Diamond, so it trades roughly 15 points of knowledge benchmark for the saving. Paying 8–18x more than Luna or V4-Flash moves you 5–7 GPQA points. Amazon Nova rows use the 2024 Nova technical report’s GPQA methodology and are not comparable with the reasoning-era scores.
Sources: developers.openai.com · openrouter.ai · openrouter.ai · openrouter.ai · the-decoder.com · datacamp.com · ai.google.dev · arxiv.org · layerlens.ai · api-docs.deepseek.com · huggingface.co · huggingface.co · mistral.ai · huggingface.co · console.groq.com · huggingface.co · assets.amazon.science
Larger vs smaller tier on GPQA Diamond (and MMLU for Nova)
%| Pair (documented distillations and same-family comparisons) | Larger / teacher tier % | Smaller tier (distilled or same-family) % |
|---|---|---|
| GPT-5.6 Sol → Luna | 92.4 | 87 |
| GPT-5.4 → GPT-5.4 mini | 93 | 88 |
| o3 → o4-mini | 83.3 | 81.4 |
| Gemini 2.5 Pro → 2.5 Flash | 86.4 | 82.8 |
| DeepSeek V4-Pro vs V4-Flash | 90.1 | 88.1 |
| DeepSeek-R1 → R1-Distill-Llama-70B | 71.5 | 65.2 |
| DeepSeek-R1 → R1-Distill-Qwen-32B | 71.5 | 62.1 |
| DeepSeek-R1 → R1-Distill-Qwen-7B | 71.5 | 49.1 |
| DeepSeek-R1 → R1-Distill-Qwen-1.5B | 71.5 | 33.8 |
| Gemma 4 31B vs Gemma 4 E4B | 84.3 | 58.6 |
| Nova Pro vs Nova Lite | 85.9 | 80.5 |
Inside a single generation the bars are nearly the same height. Across a large parameter gap they are not: R1 → R1-Distill-Qwen-1.5B loses more than half the teacher's GPQA, and Gemma 4 31B vs E4B loses 30%. Only the DeepSeek-R1 and Gemini 2.5 pairs are vendor-documented distillations; DeepSeek V4-Pro vs V4-Flash, Gemma 4 and Nova are same-family comparisons — V4-Flash is documented as a consolidation of V4 domain experts via on-policy distillation, not a compression of V4-Pro. Nova uses MMLU rather than GPQA because that is what the Amazon technical report publishes.
Sources: openrouter.ai · the-decoder.com · datacamp.com · arxiv.org · huggingface.co · huggingface.co · ai.google.dev · assets.amazon.science
Cost per 1,000,000 requests (400 input + 100 output tokens each)
USD| Model | USD per 1M requests USD |
|---|---|
| Amazon Nova Micro | 28 |
| qwen-turbo | 40 |
| Amazon Nova Lite | 48 |
| GPT-5 nano | 60 |
| gpt-oss-20b | 60 |
| Ministral 3 14B Instruct | 100 |
| qwen3.8-flash | 107 |
| GPT-4o mini | 120 |
| Mistral Small 4 | 120 |
| gpt-oss-120b | 120 |
| GPT-5.6 Luna | 200 |
| GPT-5 mini | 300 |
| DeepSeek-V4-Flash | 308 |
| Gemini 2.5 Flash | 370 |
| Gemini 3.5 Flash-Lite | 370 |
| Amazon Nova 2 Lite | 370 |
| Llama 3.3 70B Instruct | 520 |
| Gemini 3.8 Flash | 675 |
| GPT-5.4 mini | 750 |
| o4-mini | 880 |
| Claude Haiku 4.5 | 900 |
| DeepSeek-V4-Pro | 924 |
| Claude Sonnet 5 | 1,800 |
| Gemini 3.1 Pro (Preview) | 2,000 |
| GPT-5.6 Terra | 2,000 |
| GPT-5.6 Sol | 3,600 |
| Claude Opus 5 | 4,500 |
| Claude Fable 5.1 | 9,000 |
| GPT-6 Astra | 9,000 |
A classification or extraction workload at a million requests a month costs $28 on Amazon Nova Micro, $200 on GPT-5.6 Luna, $900 on Claude Haiku 4.5, $1,800 on Claude Sonnet 5 and $9,000 on the two most expensive tiers, Claude Fable 5.1 and GPT-6 Astra (both $10/$50) — a 320x spread across the same shortlist. Batch APIs cut these by 50% at both Anthropic and OpenAI, and prompt caching cuts the input component by up to 90% (97.5% on Claude Fable 5.1), which is usually a bigger lever than moving one tier down.
Sources: developers.openai.com · platform.claude.com · ai.google.dev · api-docs.deepseek.com · mistral.ai · alibabacloud.com · console.groq.com · together.ai · pricepertoken.com · pricepertoken.com · pricepertoken.com
Quality retention: smaller tier score as a percentage of the larger tier
%| Pair (documented distillations and same-family comparisons) | Knowledge benchmarks (GPQA-D / MMLU / MMLU-Pro) % | Coding / agentic benchmarks % |
|---|---|---|
| GPT-5.6 Sol → Terra | 95.7 | — |
| GPT-5.6 Sol → Luna | 94.2 | — |
| GPT-5.4 → GPT-5.4 mini | 94.6 | — |
| GPT-5.4 → GPT-5.4 nano | 89 | — |
| o3 → o4-mini | 97.7 | — |
| Gemini 2.5 Pro → 2.5 Flash | 95.8 | — |
| DeepSeek V4-Pro vs V4-Flash | 97.8 | — |
| DeepSeek-R1 → R1-Distill-Llama-70B | 91.2 | — |
| DeepSeek-R1 → R1-Distill-Qwen-32B | 86.9 | — |
| DeepSeek-R1 → R1-Distill-Qwen-14B | 82.7 | — |
| DeepSeek-R1 → R1-Distill-Llama-8B | 68.5 | — |
| DeepSeek-R1 → R1-Distill-Qwen-7B | 68.7 | — |
| DeepSeek-R1 → R1-Distill-Qwen-1.5B | 47.3 | — |
| Gemma 4 31B vs Gemma 4 E4B | 69.5 | — |
| Llama 4 Maverick → Scout | 92.3 | — |
| Nova Pro vs Nova Lite | 93.7 | — |
| Nova Pro vs Nova Micro | 90.3 | — |
| Claude Opus 5 → Sonnet 5 | — | 88.8 |
| Claude Opus 5 → Haiku 4.5 | — | 76.4 |
| Llama 4 Maverick → Scout (coding) | — | 75.6 |
| DeepSeek V4-Pro vs V4-Flash (coding) | — | 98 |
| Gemini 2.5 Pro → 2.5 Flash (coding) | — | 89.7 |
Read the two series against each other. Within one generation, knowledge retention clusters at 94–98%. Coding and agentic retention is bimodal: DeepSeek V4-Flash keeps 98% of V4-Pro on SWE-bench Verified, but Claude Haiku 4.5 keeps only 76% of Opus 5 and Llama 4 Scout only 76% of Maverick on LiveCodeBench. Only the DeepSeek-R1, Gemini 2.5 and Llama 4 rows are vendor-documented distillations; DeepSeek V4-Pro vs V4-Flash, Gemma 4 and Nova are same-family size comparisons (V4-Flash is documented as a consolidation of V4 domain experts, not a compression of V4-Pro). That spread, not the price sheet, is what should decide whether a workload can be moved down a tier.
Sources: huggingface.co · huggingface.co · arxiv.org · morphllm.com · datanorth.ai · github.com · ai.google.dev · assets.amazon.science · openrouter.ai · the-decoder.com
Enterprise inference price index, 2026
USD| Date | USD per million tokens USD |
|---|---|
| 2026-05-31 | 2.04 |
| 2026-07-25 | 1.45 |
| 2026-08-07 | 1.17 |
A 43% fall in ten weeks. Jefferies attributes it to OpenAI cutting GPT-5.6 rates by up to 80%, Anthropic matching prior frontier performance at half the price with Claude Opus 5, and Chinese open-weight models such as DeepSeek V4-Flash-0731 competing at roughly $0.03 per task. The 7 August figure is the midpoint of the reported $1.16–$1.18 low.
Tables · 5 tables
Master comparison: 74 models a buyer could shortlist in September 2026
74 rows| Model | Vendor | Role | Distilled? | Teacher | Input USD/MTok | Output USD/MTok | GPQA-D % | MMLU-Pro % | Coding | Context K tokens | Licence |
|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT-6 Astra | OpenAI | Frontier teacher | No | n/a | 10 | 50 | — | — | — | 1,050 | Proprietary API developers.openai.com |
| GPT-5.6 Sol | OpenAI | Frontier teacher | No | n/a | 4 | 20 | 92.4 | — | — | 1,050 | Proprietary API openrouter.ai |
| GPT-5.6 Terra | OpenAI | Small sibling | Undisclosed | undisclosed | 2 | 12 | 88.4 | — | — | 1,050 | Proprietary API openrouter.ai |
| GPT-5.6 Luna | OpenAI | Small sibling | Undisclosed | undisclosed | 0.2 | 1.2 | 87 | — | — | 1,050 | Proprietary API openrouter.ai |
| GPT-5.5 | OpenAI | Frontier teacher | No | n/a | 5 | 30 | — | — | — | — | Proprietary API developers.openai.com |
| GPT-5.4 | OpenAI | Frontier teacher | No | n/a | 2.5 | 15 | 93 | — | 57.7 (SWE-bench Pro) | 1,050 | Proprietary API the-decoder.com |
| GPT-5.4 mini | OpenAI | Small sibling | Undisclosed | undisclosed | 0.75 | 4.5 | 88 | — | 54.4 (SWE-bench Pro) | 400 | Proprietary API the-decoder.com |
| GPT-5.4 nano | OpenAI | Small sibling | Undisclosed | undisclosed | 0.2 | 1.25 | 82.8 | — | 52.4 (SWE-bench Pro) | 400 | Proprietary API the-decoder.com |
| GPT-5 | OpenAI | Frontier teacher | No | n/a | 1.25 | 10 | — | — | 74.9 (SWE-bench Verified) | 400 | Proprietary API arxiv.org |
| GPT-5 mini | OpenAI | Small sibling | Undisclosed | undisclosed | 0.25 | 2 | 80.3 | — | 45.7 (SWE-bench Pro) | 400 | Proprietary API openrouter.ai |
| GPT-5 nano | OpenAI | Small sibling | Undisclosed | undisclosed | 0.05 | 0.4 | 70.9 | — | — | 400 | Proprietary API openrouter.ai |
| GPT-4o | OpenAI | Frontier teacher | No | n/a | 2.5 | 10 | — | — | — | 128 | Proprietary API openrouter.ai |
| GPT-4o mini | OpenAI | Small sibling | Undisclosed | undisclosed | 0.15 | 0.6 | — | — | 87.2 (HumanEval) | 128 | Proprietary API openai.com |
| o3 | OpenAI | Frontier teacher | No | n/a | 2 | 8 | 83.3 | — | 69.1 (SWE-bench Verified) | 200 | Proprietary API datacamp.com |
| o4-mini | OpenAI | Small sibling | Undisclosed | undisclosed | 1.1 | 4.4 | 81.4 | — | 68.1 (SWE-bench Verified) | 200 | Proprietary API datacamp.com |
| gpt-oss-120b | OpenAI (open weights) | Open weights | Undisclosed | undisclosed | 0.15 | 0.6 | — | — | — | 128 | Apache 2.0 huggingface.co |
| gpt-oss-20b | OpenAI (open weights) | Open weights | Undisclosed | undisclosed | 0.075 | 0.3 | 58.6 | — | 53.2 (SWE-bench Verified) | 128 | Apache 2.0 huggingface.co |
| Claude Fable 5.1 | Anthropic | Frontier teacher | No | n/a | 10 | 50 | — | — | — | 1,000 | Proprietary API platform.claude.com |
| Claude Opus 5 | Anthropic | Frontier teacher | No | n/a | 5 | 25 | — | — | 96 (SWE-bench Verified) | 1,000 | Proprietary API datanorth.ai |
| Claude Sonnet 5 | Anthropic | Small sibling | Undisclosed | undisclosed | 2 | 10 | — | — | 85.2 (SWE-bench Verified) | 1,000 | Proprietary API morphllm.com |
| Claude Sonnet 4.6 | Anthropic | Small sibling | Undisclosed | undisclosed | 3 | 15 | — | — | — | 1,000 | Proprietary API platform.claude.com |
| Claude Haiku 4.5 | Anthropic | Small sibling | Undisclosed | undisclosed | 1 | 5 | — | — | 73.3 (SWE-bench Verified) | 200 | Proprietary API anthropic.com |
| Claude Haiku 3.5 | Anthropic | Small sibling | Undisclosed | undisclosed | 0.8 | 4 | — | — | — | 200 | Proprietary API platform.claude.com |
| Gemini 3.1 Pro (Preview) | Frontier teacher | No | n/a | 2 | 12 | 94.3 | 92.6 | 80.6 (SWE-bench Verified) | 1,000 | Proprietary API deepmind.google | |
| Gemini 3.8 Flash | Small sibling | Undisclosed | undisclosed | 0.75 | 3.75 | — | — | 73.7 (DeepSWE v1.1) | 1,000 | Proprietary API deepmind.google | |
| Gemini 3.5 Flash | Small sibling | Undisclosed | undisclosed | 1.5 | 9 | — | — | — | 1,000 | Proprietary API ai.google.dev | |
| Gemini 3.5 Flash-Lite | Small sibling | Undisclosed | undisclosed | 0.3 | 2.5 | — | — | — | 1,000 | Proprietary API ai.google.dev | |
| Gemini 3.1 Flash-Lite | Small sibling | Undisclosed | undisclosed | 0.25 | 1.5 | 72.2 | 83 | — | 1,000 | Proprietary API layerlens.ai | |
| Gemini 2.5 Pro | Frontier teacher | No | n/a | 1.25 | 10 | 86.4 | — | 67.2 (SWE-bench Verified) | 1,000 | Proprietary API arxiv.org | |
| Gemini 2.5 Flash | Documented distillation | Yes (documented) | Gemini 2.5 Pro (k-sparse logit distillation) | 0.3 | 2.5 | 82.8 | — | 60.3 (SWE-bench Verified) | 1,000 | Proprietary API arxiv.org | |
| Gemini 2.5 Flash-Lite | Documented distillation | Yes (documented) | Gemini 2.5 Pro (k-sparse logit distillation) | 0.1 | 0.4 | — | — | — | 1,000 | Proprietary API arxiv.org | |
| Gemma 4 31B | Open weights | Undisclosed | undisclosed | — | — | 84.3 | 85.2 | 80 (LiveCodeBench v6) | 256 | Apache 2.0 ai.google.dev | |
| Gemma 4 26B A4B (MoE) | Open weights | Undisclosed | undisclosed | — | — | 82.3 | 82.6 | 77.1 (LiveCodeBench v6) | 256 | Apache 2.0 ai.google.dev | |
| Gemma 4 12B Unified | Open weights | Undisclosed | undisclosed | — | — | 78.8 | 77.2 | 72 (LiveCodeBench v6) | 256 | Apache 2.0 ai.google.dev | |
| Gemma 4 E4B | Open weights | Undisclosed | undisclosed | — | — | 58.6 | 69.4 | 52 (LiveCodeBench v6) | 128 | Apache 2.0 ai.google.dev | |
| Gemma 4 E2B | Open weights | Undisclosed | undisclosed | — | — | 43.4 | 60 | 44 (LiveCodeBench v6) | 128 | Apache 2.0 ai.google.dev | |
| Gemma 3 27B IT | Open weights | Undisclosed | undisclosed | — | — | 24.3 | — | 48.8 (HumanEval) | 128 | Gemma Terms of Use (custom) huggingface.co | |
| Gemma 3 4B IT | Open weights | Undisclosed | undisclosed | — | — | 15 | — | 36 (HumanEval) | 128 | Gemma Terms of Use (custom) huggingface.co | |
| Llama 4 Maverick | Meta | Documented distillation | Yes (documented) | Llama 4 Behemoth (codistillation) | — | — | 69.8 | 80.5 | 43.4 (LiveCodeBench) | 1,000 | Llama 4 Community License (custom commercial) github.com |
| Llama 4 Scout | Meta | Documented distillation | Yes (documented) | Llama 4 Behemoth (codistillation) | — | — | 57.2 | 74.3 | 32.8 (LiveCodeBench) | 10,000 | Llama 4 Community License (custom commercial) github.com |
| Llama 3.3 70B Instruct | Meta | Open weights | Undisclosed | undisclosed | 1.04 | 1.04 | 50.5 | 68.9 | 88.4 (HumanEval) | 128 | Llama 3.3 Community License (custom commercial) github.com |
| Llama 3.2 3B Instruct | Meta | Documented distillation | Yes (documented) | Llama 3.1 8B and 70B (token-level logit distillation after pruning) | — | — | 32.8 | — | — | 128 | Llama 3.2 Community License (custom commercial) github.com |
| Llama 3.2 1B Instruct | Meta | Documented distillation | Yes (documented) | Llama 3.1 8B and 70B (token-level logit distillation after pruning) | — | — | 27.2 | — | — | 128 | Llama 3.2 Community License (custom commercial) github.com |
| DeepSeek-V4-Pro | DeepSeek | Frontier teacher | No | n/a | 1.32 | 3.96 | 90.1 | 87.5 | 80.6 (SWE-bench Verified) | 1,000 | MIT huggingface.co |
| DeepSeek-V4-Flash | DeepSeek | Documented distillation | Yes (documented) | DeepSeek V4 domain experts (on-policy distillation consolidation) | 0.44 | 1.32 | 88.1 | 86.2 | 79 (SWE-bench Verified) | 1,000 | MIT huggingface.co |
| DeepSeek-V3.2 | DeepSeek | Open weights | Yes (documented) | DeepSeek specialist models (specialist distillation into the generalist) | — | — | 82.4 | 85 | 73.1 (SWE-bench Verified) | 128 | MIT arxiv.org |
| DeepSeek-R1 (0528) | DeepSeek | Frontier teacher | No | n/a | — | — | 81 | 85 | 73.3 (LiveCodeBench) | 128 | MIT huggingface.co |
| DeepSeek-R1-Distill-Llama-70B | DeepSeek / Meta base | Documented distillation | Yes (documented) | DeepSeek-R1 | — | — | 65.2 | — | 57.5 (LiveCodeBench) | 128 | MIT (weights) over Llama 3.3 Community License base huggingface.co |
| DeepSeek-R1-Distill-Qwen-32B | DeepSeek / Qwen base | Documented distillation | Yes (documented) | DeepSeek-R1 | — | — | 62.1 | — | 57.2 (LiveCodeBench) | 128 | MIT (weights), Qwen2.5-32B base under Apache 2.0 huggingface.co |
| DeepSeek-R1-Distill-Qwen-14B | DeepSeek / Qwen base | Documented distillation | Yes (documented) | DeepSeek-R1 | — | — | 59.1 | — | 53.1 (LiveCodeBench) | 128 | MIT (weights), Qwen2.5-14B base under Apache 2.0 huggingface.co |
| DeepSeek-R1-Distill-Llama-8B | DeepSeek / Meta base | Documented distillation | Yes (documented) | DeepSeek-R1 | — | — | 49 | — | 39.6 (LiveCodeBench) | 128 | MIT (weights) over Llama 3.1 Community License base huggingface.co |
| DeepSeek-R1-Distill-Qwen-7B | DeepSeek / Qwen base | Documented distillation | Yes (documented) | DeepSeek-R1 | — | — | 49.1 | — | 37.6 (LiveCodeBench) | 128 | MIT (weights), Qwen2.5-Math-7B base under Apache 2.0 huggingface.co |
| DeepSeek-R1-Distill-Qwen-1.5B | DeepSeek / Qwen base | Documented distillation | Yes (documented) | DeepSeek-R1 | — | — | 33.8 | — | 16.9 (LiveCodeBench) | 128 | MIT (weights), Qwen2.5-Math-1.5B base under Apache 2.0 huggingface.co |
| Qwen3.8-27B | Alibaba | Open weights | Undisclosed | undisclosed | 0.5 | 3 | 89.2 | — | 61.7 (SWE-bench Pro) | 262 | Apache 2.0 huggingface.co |
| qwen3.8-max | Alibaba | Frontier teacher | No | n/a | 2 | 6 | — | — | — | — | Proprietary API alibabacloud.com |
| qwen3.8-flash | Alibaba | Small sibling | Undisclosed | undisclosed | 0.15 | 0.47 | — | — | — | — | Proprietary API alibabacloud.com |
| qwen-turbo | Alibaba | Small sibling | Undisclosed | undisclosed | 0.05 | 0.2 | — | — | — | — | Proprietary API alibabacloud.com |
| Qwen3-4B-Instruct-2507 | Alibaba | Documented distillation | Yes (documented) | Qwen3-32B / Qwen3-235B-A22B (off-policy + on-policy strong-to-weak distillation) | — | — | 62 | 69.6 | 35.1 (LiveCodeBench v6) | 262 | Apache 2.0 huggingface.co |
| qwen3-8b (hosted) | Alibaba | Documented distillation | Yes (documented) | Qwen3-32B / Qwen3-235B-A22B (strong-to-weak distillation) | 0.18 | 0.7 | — | — | — | 128 | Apache 2.0 arxiv.org |
| Phi-4 (14B) | Microsoft | Open weights | Undisclosed | undisclosed | — | — | 56.1 | 70.4 | 82.6 (HumanEval) | 16 | MIT huggingface.co |
| Phi-4-mini-instruct (3.8B) | Microsoft | Open weights | Undisclosed | undisclosed | — | — | 25.2 | 52.8 | — | 128 | MIT huggingface.co |
| Mistral Medium 3.5 | Mistral AI | Frontier teacher | No | n/a | 1.5 | 7.5 | — | — | — | — | Modified MIT docs.mistral.ai |
| Mistral Large 3 | Mistral AI | Open weights | Undisclosed | undisclosed | 0.5 | 1.5 | — | — | — | — | Apache 2.0 docs.mistral.ai |
| Mistral Small 4 | Mistral AI | Small sibling | Undisclosed | undisclosed | 0.15 | 0.6 | 71.2 | 78 | — | — | Apache 2.0 openrouter.ai |
| Ministral 3 14B Instruct | Mistral AI | Open weights | Undisclosed | undisclosed | 0.2 | 0.2 | 71.2 | — | 64.6 (LiveCodeBench) | 256 | Apache 2.0 huggingface.co |
| Ministral 3 8B | Mistral AI | Open weights | Undisclosed | undisclosed | 0.15 | 0.15 | — | — | — | 256 | Apache 2.0 docs.mistral.ai |
| Ministral 3 3B | Mistral AI | Open weights | Undisclosed | undisclosed | 0.1 | 0.1 | — | — | — | 256 | Apache 2.0 docs.mistral.ai |
| Amazon Nova Premier | Amazon | Frontier teacher | No | n/a | — | — | — | — | — | 1,000 | Proprietary API docs.aws.amazon.com |
| Amazon Nova Pro | Amazon | Frontier teacher | Undisclosed (Bedrock distillation student) | undisclosed | 0.8 | 3.2 | 46.9 | — | — | 300 | Proprietary API assets.amazon.science |
| Amazon Nova Lite | Amazon | Small tier | Undisclosed (Bedrock distillation student) | undisclosed | 0.06 | 0.24 | 42 | — | — | 300 | Proprietary API assets.amazon.science |
| Amazon Nova Micro | Amazon | Small tier | Undisclosed (Bedrock distillation student) | undisclosed | 0.035 | 0.14 | 40 | — | — | 128 | Proprietary API assets.amazon.science |
| Amazon Nova 2 Lite | Amazon | Small sibling | Undisclosed | undisclosed | 0.3 | 2.5 | — | — | — | 1,000 | Proprietary API docs.aws.amazon.com |
| SmolLM3-3B | Hugging Face | Open weights | Undisclosed | n/a | — | — | 41.7 | — | 30.48 (HumanEval+) | 128 | Apache 2.0 huggingface.co |
| grok-4.6 | xAI | Frontier teacher | No | n/a | 2 | 6 | — | — | — | 200 | Proprietary API docs.x.ai |
Prices are vendor list rates in USD per million tokens, before batch (typically -50%) or cache discounts. Open-weight rows with no price are self-host only in this dataset. DeepSeek prices are peak-hour; off-peak is half. Gemini 3.1 Pro and 2.5 Pro input prices double above 200K input tokens. Gemini 3.8 Flash's $0.75/$3.75 is promotional through 2026-12-31; it lists at $1.50/$7.50 from 2027-01-01. Qwen3.8-27B is shown at Alibaba Model Studio's list rate; Groq hosts the same weights at $0.80/$4.00. Gemini 3.7 Flash shipped in August 2026 but is not included in this shortlist.
Sources: developers.openai.com · platform.claude.com · ai.google.dev · api-docs.deepseek.com · mistral.ai · alibabacloud.com · together.ai · console.groq.com · ai.google.dev · github.com · huggingface.co · assets.amazon.science
Price vs quality: GPQA Diamond points per dollar of blended token price
24 rows| Model | Vendor | Tier | GPQA Diamond % | Blended price USD/MTok | GPQA points per $ pts/USD | Cost / 1M requests USD |
|---|---|---|---|---|---|---|
| Amazon Nova Micro | Amazon | Small tier | 40 | 0.061 | 653 | 28 assets.amazon.science |
| GPT-5 nano | OpenAI | Small sibling | 70.9 | 0.138 | 516 | 60 openrouter.ai |
| gpt-oss-20b | OpenAI (open weights) | Open weights | 58.6 | 0.131 | 447 | 60 huggingface.co |
| Amazon Nova Lite | Amazon | Small tier | 42 | 0.105 | 400 | 48 assets.amazon.science |
| Ministral 3 14B Instruct | Mistral AI | Open weights | 71.2 | 0.2 | 356 | 100 huggingface.co |
| Mistral Small 4 | Mistral AI | Small sibling | 71.2 | 0.263 | 271 | 120 openrouter.ai |
| GPT-5.6 Luna | OpenAI | Small sibling | 87 | 0.45 | 193 | 200 openrouter.ai |
| GPT-5.4 nano | OpenAI | Small sibling | 82.8 | 0.463 | 179 | 205 the-decoder.com |
| Gemini 3.1 Flash-Lite | Small sibling | 72.2 | 0.563 | 128 | 250 layerlens.ai | |
| DeepSeek-V4-Flash | DeepSeek | Distilled | 88.1 | 0.66 | 134 | 308 huggingface.co |
| GPT-5 mini | OpenAI | Small sibling | 80.3 | 0.688 | 117 | 300 openrouter.ai |
| Gemini 2.5 Flash | Distilled | 82.8 | 0.85 | 97.4 | 370 arxiv.org | |
| Qwen3.8-27B | Alibaba | Open weights | 89.2 | 1.13 | 79.3 | 500 huggingface.co |
| GPT-5.4 mini | OpenAI | Small sibling | 88 | 1.69 | 52.1 | 750 the-decoder.com |
| Llama 3.3 70B Instruct | Meta | Open weights | 50.5 | 1.04 | 48.6 | 520 github.com |
| DeepSeek-V4-Pro | DeepSeek | Teacher | 90.1 | 1.98 | 45.5 | 924 huggingface.co |
| o4-mini | OpenAI | Small sibling | 81.4 | 1.93 | 42.3 | 880 openrouter.ai |
| Amazon Nova Pro | Amazon | Teacher | 46.9 | 1.4 | 33.5 | 640 assets.amazon.science |
| o3 | OpenAI | Teacher | 83.3 | 3.5 | 23.8 | 1,600 datacamp.com |
| Gemini 2.5 Pro | Teacher | 86.4 | 3.44 | 25.1 | 1,500 arxiv.org | |
| Gemini 3.1 Pro (Preview) | Teacher | 94.3 | 4.5 | 21 | 2,000 deepmind.google | |
| GPT-5.6 Terra | OpenAI | Small sibling | 88.4 | 4.5 | 19.6 | 2,000 openrouter.ai |
| GPT-5.4 | OpenAI | Teacher | 93 | 5.63 | 16.5 | 2,500 the-decoder.com |
| GPT-5.6 Sol | OpenAI | Teacher | 92.4 | 8 | 11.6 | 3,600 openrouter.ai |
GPT-5 nano tops this ranking on raw efficiency but scores only 70.9 GPQA; GPT-5.6 Luna is the highest-scoring model in the top three, which is why it dominates most 2026 routing configurations. Amazon Nova figures use MMLU-era GPQA methodology from the Nova technical report and are not directly comparable with the 2026 reasoning-model scores.
Sources: developers.openai.com · openrouter.ai · api-docs.deepseek.com · huggingface.co · ai.google.dev · arxiv.org · assets.amazon.science · console.groq.com
Measured latency and throughput (Artificial Analysis, September 2026)
12 rows| Model (setting) | Time to first token s | Output speed tokens/s | AA Intelligence Index | Blended price USD/MTok |
|---|---|---|---|---|
| Gemini 2.5 Flash-Lite (non-reasoning) | 0.3 | — | — | 0.18 artificialanalysis.ai |
| gpt-oss-120b (high) | 0.85 | 151 | 24 | 0.2 artificialanalysis.ai |
| Llama 4 Maverick | 0.92 | 82 | 14 | 0.31 artificialanalysis.ai |
| DeepSeek V4-Flash 0731 (reasoning, max) | 1.19 | 140 | 52 | 0.23 artificialanalysis.ai |
| GPT-5.6 Luna (low) | 1.7 | 109 | 34 | 0.17 artificialanalysis.ai |
| DeepSeek V4-Pro 0813 (reasoning, max) | 1.9 | 60.2 | 53 | 0.69 artificialanalysis.ai |
| Gemini 3.5 Flash-Lite | 6.48 | 391 | 37 | 0.33 artificialanalysis.ai |
| GPT-5.6 Luna (high) | 19.9 | 123 | 47 | 0.17 artificialanalysis.ai |
| Claude Haiku 4.5 (reasoning) | 19.9 | 90 | 30 | 0.77 artificialanalysis.ai |
| Claude Opus 5 (adaptive, max effort) | 77.3 | 57 | 63 | 3.85 artificialanalysis.ai |
| Claude Sonnet 5 (adaptive, max effort) | 178 | 78 | 55 | 1.54 artificialanalysis.ai |
| Amazon Nova 2 Lite (non-reasoning) | — | 149 | 12 | 0.85 pricepertoken.com |
Gemini 3.5 Flash-Lite is the throughput leader at 391 tokens/s. GPT-5.6 Luna is the only model here that spans both ends of the latency range purely through its effort parameter (1.70s to 19.87s), which makes it unusually easy to run interactive and batch traffic on one model ID. Claude Sonnet 5’s 177.77s figure is max-effort adaptive thinking, not a typical production setting.
Sources: artificialanalysis.ai · artificialanalysis.ai · artificialanalysis.ai · artificialanalysis.ai · artificialanalysis.ai · artificialanalysis.ai
Licence and deployment: what you are allowed to do with each family
14 rows| Family | Licence | Weights | Self-host / air-gap | Distil outputs? | Data residency options | Buyer note |
|---|---|---|---|---|---|---|
| OpenAI GPT-5.x / GPT-6 / o-series | Proprietary API | Closed | No | No — Services Agreement bars using Output to develop competing models | Azure/Foundry regions | Deepest price ladder in the market and 1.05M context on every 5.6 tier. openai.com |
| OpenAI gpt-oss 20b / 120b | Apache 2.0 | Open | Yes (120b on one 80GB GPU; 20b in 16GB) | Yes | Anywhere | The escape hatch inside the OpenAI ecosystem. Served by Groq at ~500–1000 tok/s. huggingface.co |
| Anthropic Claude (Fable / Opus / Sonnet / Haiku) | Proprietary API | Closed | No | No — Commercial Terms D.4 bars building a competing product or training competing AI models | inference_geo:"us" at a 1.1x multiplier; Bedrock/Vertex regional endpoints at +10% | Best published SWE-bench Verified (Opus 5, 96.0). 1M context on Fable/Opus/Sonnet, 200K on Haiku 4.5. anthropic.com |
| Google Gemini 2.5 / 3.x | Proprietary API | Closed | No | No — Gemini API Additional Terms restrict competitive model development | Vertex AI regions | Only closed vendor that has publicly documented distilling its own small tier (Gemini 2.5 report). arxiv.org |
| Google Gemma 4 | Apache 2.0 | Open | Yes (2.3B–31B) | Yes | Anywhere | MMLU-Pro 85.2 at 31B under Apache 2.0 — the strongest permissively-licensed model you can put on one node. ai.google.dev |
| Google Gemma 3 | Gemma Terms of Use (custom, use restrictions apply) | Open | Yes | Yes, subject to the Gemma prohibited-use policy | Anywhere | Superseded by Gemma 4 on both quality and licence terms. huggingface.co |
| Meta Llama 3.x / 4 | Llama Community License (custom commercial; 700M MAU clause) | Open | Yes | Yes, with Llama attribution and naming obligations on derivatives | Anywhere | Llama 3.2 1B/3B are explicitly documented distillations; Meta's Llama 4 launch post describes Scout/Maverick as codistilled from the unreleased Behemoth (the model card itself does not mention it). ai.meta.com |
| DeepSeek R1 / V3.2 / V4 | MIT | Open | Yes (V4-Pro is a 1.6T MoE — non-trivial) | Yes — the R1 release shipped six distilled students itself | Anywhere; first-party API is PRC-hosted | Highest open-weight GPQA Diamond here (90.1). Many enterprises self-host rather than use the PRC-hosted API. huggingface.co |
| Alibaba Qwen3 / Qwen3.8 (open weights) | Apache 2.0 | Open | Yes | Yes | Anywhere; Model Studio API is PRC/Singapore | Qwen3 technical report documents strong-to-weak distillation for the 0.6B–14B dense sizes and 30B-A3B. huggingface.co |
| Microsoft Phi-4 / Phi-4-mini | MIT | Open | Yes | Yes | Anywhere | Least legally encumbered small models in this dataset. Phi-4’s 16K context is the catch. huggingface.co |
| Mistral Small 4 / Ministral 3 / Large 3 | Apache 2.0 | Open | Yes | Yes | EU-headquartered vendor; EU hosting available | The default answer when the requirement is EU sovereignty plus a permissive licence. docs.mistral.ai |
| Mistral Medium 3.5 | Modified MIT | Open | Yes | Yes, subject to the modified terms | EU hosting available | Frontier-class tier of the Mistral line; read the modification before assuming MIT. docs.mistral.ai |
| Amazon Nova / Nova 2 | Proprietary API | Closed | No | Only through Bedrock Model Distillation, into another Amazon-supported student | AWS regions incl. GovCloud (US-West) | The only vendor that documents its teacher/student graph in product docs: Premier → Pro/Lite/Micro, Pro → Lite/Micro. docs.aws.amazon.com |
| Hugging Face SmolLM3 | Apache 2.0 | Open (plus data mixture and training configs) | Yes | Yes | Anywhere | Fully reproducible supply chain — relevant where an auditor asks what the model was trained on. huggingface.co |
Anti-distillation clauses bind the enterprise account holder, not just individual developers, and survive termination in most of these agreements. If your roadmap includes training an in-house model on teacher outputs, pick an MIT or Apache 2.0 teacher up front rather than seeking a waiver later.
Sources: anthropic.com · openai.com · huggingface.co · ai.google.dev · github.com · huggingface.co · huggingface.co · huggingface.co · docs.mistral.ai · docs.aws.amazon.com · huggingface.co
Quality retention by pair (documented distillations and same-family size comparisons)
22 rows| Teacher → student | Relationship | Benchmark | Teacher % | Student % | Retention % | Student price / teacher price (input) x |
|---|---|---|---|---|---|---|
| GPT-5.6 Sol → Terra | Same-generation tier comparison (method undisclosed) | GPQA Diamond | 92.4 | 88.4 | 95.7 | 0.5 openrouter.ai |
| GPT-5.6 Sol → Luna | Same-generation tier comparison (method undisclosed); third-party benchmark, 94.2–97.3% across providers | GPQA Diamond | 92.4 | 87 | 94.2 | 0.05 openrouter.ai |
| GPT-5.4 → GPT-5.4 mini | Same-generation tier comparison (method undisclosed) | GPQA Diamond | 93 | 88 | 94.6 | 0.3 the-decoder.com |
| GPT-5.4 → GPT-5.4 nano | Same-generation tier comparison (method undisclosed) | GPQA Diamond | 93 | 82.8 | 89 | 0.08 the-decoder.com |
| o3 → o4-mini | Same-generation tier comparison (method undisclosed) | GPQA Diamond | 83.3 | 81.4 | 97.7 | 0.55 datacamp.com |
| Gemini 2.5 Pro → 2.5 Flash | Documented distillation | GPQA Diamond | 86.4 | 82.8 | 95.8 | 0.24 arxiv.org |
| DeepSeek V4-Pro vs V4-Flash | Same-family price/quality comparison — V4-Flash is a consolidation of V4 domain experts via on-policy distillation, not a compression of V4-Pro | GPQA Diamond | 90.1 | 88.1 | 97.8 | 0.333 huggingface.co |
| DeepSeek-R1 → R1-Distill-Llama-70B | Documented distillation | GPQA Diamond | 71.5 | 65.2 | 91.2 | — huggingface.co |
| DeepSeek-R1 → R1-Distill-Qwen-32B | Documented distillation | GPQA Diamond | 71.5 | 62.1 | 86.9 | — huggingface.co |
| DeepSeek-R1 → R1-Distill-Qwen-14B | Documented distillation | GPQA Diamond | 71.5 | 59.1 | 82.7 | — huggingface.co |
| DeepSeek-R1 → R1-Distill-Llama-8B | Documented distillation | GPQA Diamond | 71.5 | 49 | 68.5 | — huggingface.co |
| DeepSeek-R1 → R1-Distill-Qwen-7B | Documented distillation | GPQA Diamond | 71.5 | 49.1 | 68.7 | — huggingface.co |
| DeepSeek-R1 → R1-Distill-Qwen-1.5B | Documented distillation | GPQA Diamond | 71.5 | 33.8 | 47.3 | — huggingface.co |
| Gemma 4 31B vs Gemma 4 E4B | Same-family size comparison — Google does not describe E4B as a distillation of 31B | GPQA Diamond | 84.3 | 58.6 | 69.5 | — ai.google.dev |
| Llama 4 Maverick → Scout | Documented codistillation (both from Behemoth), compared here by size | MMLU-Pro | 80.5 | 74.3 | 92.3 | — github.com |
| Nova Pro vs Nova Lite | Same-family size comparison — Bedrock supports both as distillation students; shipped provenance undisclosed | MMLU | 85.9 | 80.5 | 93.7 | 0.075 assets.amazon.science |
| Nova Pro vs Nova Micro | Same-family size comparison — Bedrock supports both as distillation students; shipped provenance undisclosed | MMLU | 85.9 | 77.6 | 90.3 | 0.044 assets.amazon.science |
| Claude Opus 5 → Sonnet 5 | Same-generation tier comparison (method undisclosed) | SWE-bench Verified | 96 | 85.2 | 88.8 | 0.4 morphllm.com |
| Claude Opus 5 → Haiku 4.5 | Cross-generation tier comparison (method undisclosed) | SWE-bench Verified | 96 | 73.3 | 76.4 | 0.2 datanorth.ai |
| Llama 4 Maverick → Scout (coding) | Documented codistillation (both from Behemoth), compared here by size | LiveCodeBench | 43.4 | 32.8 | 75.6 | — github.com |
| DeepSeek V4-Pro vs V4-Flash (coding) | Same-family price/quality comparison — see row 6 | SWE-bench Verified | 80.6 | 79 | 98 | 0.333 huggingface.co |
| Gemini 2.5 Pro → 2.5 Flash (coding) | Documented distillation | SWE-bench Verified | 67.2 | 60.3 | 89.7 | 0.24 arxiv.org |
Two patterns stand out. First, knowledge retention above 94% is now routine inside a family, and DeepSeek V4-Flash reaches 97.8% for a third of the price. Second, retention falls off a cliff for very small students on broad-knowledge benchmarks — R1-Distill-Qwen-1.5B keeps only 47% of R1’s GPQA — while the same model keeps 86% of R1’s MATH-500. Distillation buys you narrow competence cheaply and broad competence expensively. Only rows marked "Documented distillation" carry a vendor statement that the student was trained from the teacher. The remaining rows are same-family or same-generation price/quality comparisons and should not be read as training provenance.
Sources: openrouter.ai · the-decoder.com · arxiv.org · huggingface.co · huggingface.co · ai.google.dev · github.com · assets.amazon.science · morphllm.com · datanorth.ai · datacamp.com
Timeline · 21 events
GPT-4o mini launches at $0.15/$0.60
MMLU 82.0, HumanEval 87.2, MMMU 59.4 — over 60% cheaper than GPT-3.5 Turbo. Establishes the "cheap tier" as a permanent product line every vendor now copies.
Source: openai.comLlama 3.2 1B/3B ship as documented distillations
Meta states that logits from Llama 3.1 8B and 70B were used as token-level targets in pretraining, with distillation applied after pruning to recover performance. The first mainstream model card to spell out the recipe.
Source: github.comAmazon Bedrock Model Distillation announced
Turns distillation into a managed service: the customer supplies prompts, AWS generates teacher responses and fine-tunes the student. Later reaches GA with claims of up to 500% faster and 75% cheaper inference.
Source: aws.amazon.comLlama 3.3 70B released
MMLU 86.0 CoT, MMLU-Pro 68.9, HumanEval 88.4, 128K context under the Llama 3.3 Community License. Becomes the default open student base for enterprise distillation projects.
Source: github.comDeepSeek ships R1 plus six distilled students under MIT
R1-Distill-Qwen-32B beats o1-mini on AIME 2024 (72.6 vs 63.6), MATH-500 (94.3 vs 90.0) and GPQA Diamond (62.1 vs 60.0). For buyers this was the first time a free download credibly replaced a paid reasoning tier.
Source: huggingface.coLlama 4 Scout and Maverick ship as codistilled models
Meta's launch post says both are codistilled from the ~2T-parameter Llama 4 Behemoth, which was never released; the Llama 4 model card carries the benchmarks but not the provenance claim. Maverick posts MMLU-Pro 80.5 / GPQA-D 69.8; Scout adds a 10M-token context window.
Source: ai.meta.comGemini 2.5 report confirms the Flash line is distilled
Google states the smaller 2.5 models are distilled and that the teacher’s next-token distribution is approximated with a k-sparse distribution to cut storage cost. Rare public confirmation from a closed-model vendor.
Source: arxiv.orgGPT-5, GPT-5 mini and GPT-5 nano launch together
A three-tier release at $1.25/$10, $0.25/$2 and $0.05/$0.40 with 400K context on all three. Tiered families become the default shape of a frontier launch.
Source: developers.openai.comClaude Haiku 4.5 released at $1/$5
SWE-bench Verified 73.3. Anthropic positions it as matching Claude Sonnet 4 coding performance at a third of the cost and more than twice the speed.
Source: anthropic.comMinistral 3 (3B/8B/14B) ships Apache 2.0 with 256K context
The 14B posts GPQA Diamond 71.2, AIME25 85.0 and MATH 90.4 — edge-class weights with no licence friction, which matters for EU buyers.
Source: huggingface.coMistral Small 4 released under Apache 2.0
MMLU-Pro 78.0, GPQA Diamond 71.2, priced at $0.15/$0.60 on the Mistral API. Becomes the reference EU-sovereign small model.
Source: openrouter.aiGemma 4 ships Apache 2.0 with frontier-class small models
The 31B dense posts MMLU-Pro 85.2 and GPQA Diamond 84.3; the 26B MoE reaches 82.3 GPQA with only 3.8B active parameters. Google moves the Gemma line from a custom licence to Apache 2.0.
Source: deepmind.googleDeepSeek-V4-Pro released under MIT
1.6T total / 49B active parameters, 1M context, MMLU-Pro 87.5, GPQA Diamond 90.1, SWE-bench Verified 80.6. The strongest openly-licensed model a buyer can self-host.
Source: huggingface.coClaude Sonnet 5 launches at $2/$10 with 1M context
SWE-bench Verified 85.2, Terminal-Bench 2.1 80.4 — close to Opus 4.8 at a fraction of the price. Introductory pricing later made permanent.
Source: anthropic.comGPT-5.6 Sol, Terra and Luna launch as one price ladder
All three carry a 1.05M-token context. GPQA Diamond 92.4 / 88.4 / 87.0 across a 20x price spread — the clearest published price-vs-quality ladder in the market.
Source: openrouter.aiClaude Opus 5 released at $5/$25
SWE-bench Verified 96.0, SWE-bench Pro 79.2, OSWorld 2.0 70.6. Anthropic positions it as near-Fable frontier quality at half the price, holding the Opus price flat.
Source: platform.claude.comOpenAI cuts GPT-5.6 Luna by 80% and Terra by 20%
Luna drops to $0.20/$1.20. This single move reset the floor for closed-model pricing and is a major driver of the mid-2026 fall in the enterprise inference index.
Source: axios.comEnterprise inference index hits a 2026 low of $1.16–$1.18/MTok
Silicon Data index cited by Jefferies, down from $2.04 on 31 May and $1.45 in late July, driven by OpenAI’s cuts, Anthropic’s Opus 5 repricing and Chinese open-weight competition.
Source: scmp.comAnthropic makes Sonnet 5 $2/$10 permanent
The scheduled 1 September increase to $3/$15 is cancelled. A rare case of a vendor withdrawing an announced price rise under competitive pressure.
Source: platform.claude.comDeepSeek raises V4 standard rates ~3–4.7x, and cache-hit input rates by up to 11x
Standard rates: V4-Flash goes from $0.14/$0.28 to $0.44/$1.32 at peak (about 3.1x input, 4.7x output); V4-Pro from $0.435/$0.87 to $1.32/$3.96 (about 3.0x / 4.6x). InfoWorld's "more than 10x" headline refers specifically to cache-hit input tokens, which rose between 52% and 1,100%. Capacity, not competition, sets the floor for the cheapest tiers.
Source: infoworld.comGemini 3.8 Flash ships at $0.75/$3.75
HLE-Verified 54.9 and 1M context per 9to5Google; Terminal-Bench 2.1 89.4 and DeepSWE v1.1 73.7 per Google's Gemini Flash model page (https://deepmind.google/models/gemini/flash/). Google's third Flash update in three months — the previous model, Gemini 3.7 Flash, shipped three weeks earlier — evidence that the cheap tier is now the fastest-moving part of the market.
Source: 9to5google.com
Glossary · 16 terms
- Knowledge distillation
- Training a small "student" model to reproduce the behaviour of a large "teacher" — either from its output text (black-box) or from its output probability distribution over tokens (white-box logit distillation).
- Teacher / student
- The large source model and the small target model in a distillation. AWS documents Nova Premier as a teacher to Nova Pro, Lite and Micro, and Nova Pro as a teacher to Lite and Micro.
- Quality retention
- The student’s benchmark score as a percentage of the teacher’s on the same benchmark. Useful for buyers because it is scale-free, but it varies enormously by benchmark: broad knowledge retains worse than narrow maths.
- Small sibling
- A cheaper tier released alongside a frontier model (mini, nano, Flash, Haiku, Luna) where the vendor has not publicly documented how it was built. Behaves like a distillation commercially whether or not it is one technically.
- k-sparse distillation
- Storing only the top-k entries of the teacher’s next-token probability distribution instead of the full vocabulary, to make logit distillation affordable at scale. Documented in the Gemini 2.5 technical report.
- Codistillation
- Training several student models jointly against a shared teacher during the teacher’s own training run. Meta used this for Llama 4 Scout and Maverick against Llama 4 Behemoth.
- Strong-to-weak distillation
- Alibaba’s two-phase Qwen3 pipeline: off-policy distillation on teacher outputs in both thinking and non-thinking modes, then on-policy distillation aligning student logits with a Qwen3-32B or 235B-A22B teacher by KL divergence.
- GPQA Diamond
- A 198-question set of graduate-level science problems written to be resistant to web search. The most commonly quoted knowledge benchmark for 2025–26 frontier models.
- MMLU-Pro
- A harder, ten-choice successor to MMLU with more reasoning-heavy questions. Scores are typically 10–20 points below MMLU for the same model, so the two are not interchangeable in a comparison table.
- SWE-bench Verified
- A 500-issue human-validated subset of SWE-bench measuring whether a model can resolve a real GitHub issue end to end. The benchmark where distilled tiers lose the most ground.
- Time to first token (TTFT)
- Latency from request to first streamed token. On reasoning models it now includes thinking time, so a "fast" model at high effort can measure slower than a "slow" model at low effort.
- Blended price
- A single price per million tokens computed as 0.75 x input + 0.25 x output, the 3:1 weighting Artificial Analysis uses. Handy for ranking, misleading for workloads with unusual input:output ratios.
- Prompt caching
- Charging a reduced rate for repeated prefix tokens. Cache reads cost 10% of base input on most Claude models and 2.5% on Fable 5.1; OpenAI caches at roughly 10% of input. Often a larger saving than switching model tiers.
- Anti-distillation clause
- Contract language barring customers from using a vendor’s outputs to train a competing model. Anthropic’s Commercial Terms D.4 and OpenAI’s Services Agreement both contain one; MIT and Apache 2.0 open-weight models do not.
- Model tiering / routing
- Sending most traffic to a cheap model and escalating only hard requests to a frontier tier. The dominant 2026 cost-control pattern and the main practical way distillation shows up on a buyer’s invoice.
- Effective parameters
- For MoE and Matformer-style models, the parameters actually activated per token (e.g. Gemma 4 26B A4B activates 3.8B of 25.2B; DeepSeek-V4-Flash activates 13B of 284B). Determines serving cost far more than total parameter count.
Sources · 78 sources
Every figure on this page comes from one of these primary sources. Compiled 4 September 2026.
- OpenAI API pricing
- OpenAI API models reference
- GPT-4o mini: advancing cost-efficient intelligence
- OpenAI GPT-5 System Card
- OpenAI Services Agreement
- GPT-5.6 Sol model page
- GPT-5.6 Terra model page
- GPT-5.6 Luna model page
- GPT-5 mini model page
- GPT-5 nano model page
- GPT-4o model page
- o4-mini model page
- OpenAI ships GPT-5.4 mini and nano
- o4-mini: tests, features, o3 comparison, benchmarks
- gpt-oss-20b model card
- Claude platform pricing
- Claude models overview
- Claude Opus 5 model page
- Introducing Claude Sonnet 5
- Introducing Claude Haiku 4.5
- Anthropic Commercial Terms of Service
- Claude Opus 5 by Anthropic: benchmarks and pricing
- Claude benchmarks 2026
- Anthropic launches Opus 5
- Gemini API pricing
- Gemini API models
- Gemini 3.1 Pro model page
- Gemini Flash model page
- Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context and Next Generation Agentic Capabilities
- Gemma 4 model card
- Gemma 4
- gemma-3-27b-it model card
- Gemini 3.8 Flash rolling out three weeks after last release
- Gemini 3.8 Flash model stats
- Gemini 3.1 Flash-Lite benchmark results
- DeepSeek API pricing
- DeepSeek-R1 model card (with distilled model evaluations)
- DeepSeek-R1-0528 model card
- DeepSeek-R1-Distill-Qwen-32B model card
- DeepSeek-V4-Pro model card
- DeepSeek-V4-Flash model card
- DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
- DeepSeek raises some V4 prices by more than 10x
- Llama 4 model card
- Llama 3.3 model card
- Llama 3.2 model card
- Qwen3 Technical Report
- Qwen3.8-27B model card
- Qwen3-4B-Instruct-2507 model card
- Alibaba Model Studio model pricing
- phi-4 model card
- Phi-4-mini-instruct model card
- SmolLM3-3B model card
- Mistral AI API pricing
- Mistral models overview
- Ministral-3-14B-Instruct-2512 model card
- Mistral Small 4 model page
- What is Amazon Nova? (teacher/student distillation matrix)
- What’s new in Amazon Nova 2
- The Amazon Nova Family of Models: Technical Report and Model Card
- Amazon Bedrock Model Distillation
- Amazon Bedrock pricing
- Nova Micro API pricing
- Nova Lite API pricing
- Nova 2 Lite API pricing
- Amazon Nova Pro: AWS Bedrock model guide, specs and pricing (2026)
- Together AI pricing
- Groq supported models and pricing
- xAI models and pricing
- GPT-5.6 Luna (low) vs Claude 4.5 Haiku (reasoning)
- Gemini 3.5 Flash-Lite vs GPT-5.6 Luna (high)
- Claude Sonnet 5 vs Claude Opus 5
- DeepSeek V4 Flash vs DeepSeek V4 Pro
- gpt-oss-120B vs Llama 4 Maverick
- Comparison of AI models across intelligence, performance and price
- Enterprise AI costs hit 2026 low driven by price wars and Chinese open-source models
- OpenAI discounts GPT-5.6 Luna and Terra
- The Llama 4 herd: the beginning of a new era of natively multimodal AI innovation