page_url,page_updated,theme,metric_id,metric,unit,year,value,value_numeric,as_of,source,source_url,confidence,note https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,swe_bench_verified_top,"SWE-bench Verified, top reported score",% of tasks resolved,2024,71.7%,71.7,Dec 2024,"OpenAI (o3 preview announcement), as reported in Stanford HAI AI Index 2025",https://hai.stanford.edu/ai-index/2025-ai-index-report/technical-performance,MEDIUM,"o3 was a preview, not generally available, when this score was announced. Stanford HAI pairs it with a 4.4% 2023 figure that was measured on the original full SWE-bench, not SWE-bench Verified (released Aug 2024); the two endpoints are different test versions. Claude 3.5 Sonnet (Oct 2024) was the best generally available model at about 49%." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,swe_bench_verified_top,"SWE-bench Verified, top reported score",% of tasks resolved,2025,80.9%,80.9,24 Nov 2025,"Anthropic (Claude Opus 4.5 launch), vendor-reported",https://www.anthropic.com/news/claude-opus-4-5,MEDIUM,"First model reported above 80%. Rivals at the time: GPT-5.1-Codex-Max 77.9%, Claude Sonnet 4.5 77.2%, Gemini 3 Pro 76.2% (vendor figures, secondary reporting)." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,swe_bench_verified_top,"SWE-bench Verified, top reported score",% of tasks resolved,2026,97.0% (to date),97.0,1 Sep 2026,"Vals.ai SWE-bench Verified leaderboard (Claude Opus 5, mini-SWE-agent harness)",https://www.vals.ai/benchmarks/swebench,MEDIUM,"Single independent harness, read via search summary. DeepSeek V4 Pro 96.4% and GPT-5.6 Sol 96.2% on the same board. BenchLM (29 Jul 2026) shows 96%; vendor figure for Claude Opus 4.8 was 88.6% (May 2026, page 964). Vals says the benchmark is saturated and no longer runs it on new releases. Stanford HAI 2026 describes the same move as about 60% to nearly 100% of the human baseline in one year. Harness differences mean 2026 is not strictly comparable with the vendor-run 2024–25 figures." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,gpqa_diamond_top,"GPQA Diamond (graduate-level science questions), top reported score",% correct,2024,87.7%,87.7,Dec 2024,OpenAI (o3 preview announcement),,MEDIUM,"Vendor-reported for a preview model; not re-verified in this session. Generally available o1 (Sep–Dec 2024) reported 78%. Stanford HAI records a +48.9 percentage-point one-year gain on GPQA, 2023 to 2024." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,gpqa_diamond_top,"GPQA Diamond (graduate-level science questions), top reported score",% correct,2025,93.8%,93.8,Dec 2025,"Google (Gemini 3 Deep Think), vendor-reported",,MEDIUM,GPT-5.2 Pro reported 93.2% and Gemini 3 Pro 91.9% in the same period (secondary reporting of vendor figures). https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,gpqa_diamond_top,"GPQA Diamond (graduate-level science questions), top reported score",% correct,2026,94.3% (to date),94.3,Feb 2026,"Google DeepMind, Gemini 3.1 Pro model card",https://deepmind.google/models/model-cards/gemini-3-1-pro/,MEDIUM,"An aggregator lists GPT-5.4 Pro at 94.4% (self-reported). Expert review suggests only about 90–95% of GPQA Diamond questions are valid, so scores in the mid-90s are at the ceiling: the benchmark is saturated." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,hle_top,"Humanity's Last Exam, top reported score",% correct,2024,~9%,9,Jan 2025 (benchmark launch; end-2024 models),"Center for AI Safety / Scale AI launch results (o1), as cited by OpenAI",https://agi.safe.ai/,MEDIUM,"HLE launched in January 2025; the best end-2024 models (o1, DeepSeek-R1) scored about 9%. Listed under 2024 because it measures 2024 models." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,hle_top,"Humanity's Last Exam, top reported score",% correct,2025,41.0% (no tools),41.0,Dec 2025,"Google (Gemini 3 Deep Think), as tabulated by R&D World",https://www.rdworldonline.com/how-gpt-5-2-stacks-up-against-gemini-3-0-and-claude-opus-4-5/,CONTESTED,"Another source gives 45.1% for the December Deep Think version. With tools, OpenAI Deep Research scored 26.6% in Feb 2025 (page 386); tool and no-tool scores are not comparable. Stanford HAI 2026 reports a roughly +30 pp one-year frontier gain on HLE." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,hle_top,"Humanity's Last Exam, top reported score",% correct,2026,54.8% (to date),54.8,17 Oct 2026,"Scale Labs HLE leaderboard (GPT-6 Astra), via BenchLM mirror",https://labs.scale.com/leaderboard/humanitys_last_exam,CONTESTED,"Scale's own page could not be opened. A mirror of a 29 Sep 2026 snapshot shows GPT-6 Astra at 60.6%; another tracker lists Claude Mythos 5 at 59% (no tools, 20 Aug 2026); Artificial Analysis's own HLE run shows Claude Fable 5 at 53.3%. Gemini 3 Deep Think reported 48.4% on 12 Feb 2026. Treat the 2026 top as about 55–60% depending on leaderboard and setting." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,arc_agi_top,"ARC-AGI, top verified score (version changes: ARC-AGI-1 in 2024, ARC-AGI-2 from 2025)","% of tasks, semi-private set",2024,75.7% (ARC-AGI-1),75.7,20 Dec 2024,"ARC Prize Foundation (OpenAI o3 preview, within $10k compute limit)",https://arcprize.org/blog/oai-o3-pub-breakthrough,MEDIUM,87.5% in a high-compute (172×) configuration. The tested model was a preview trained partly on ARC-AGI-1 public training data; the released o3 later scored about 41%. Different benchmark from 2025–26 values: do not chart as one series. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,arc_agi_top,"ARC-AGI, top verified score (version changes: ARC-AGI-1 in 2024, ARC-AGI-2 from 2025)","% of tasks, semi-private set",2025,54% (ARC-AGI-2),54,5 Dec 2025,ARC Prize Foundation verification of Poetiq (meta-system on Gemini 3 Pro),https://bdtechtalks.com/2025/12/09/poetiq-arc-agi-2-solution/,MEDIUM,At about $30.57 per task; Gemini 3 Deep Think had scored 45.1% at about $77 per task. ARC-AGI-2 launched in March 2025 because ARC-AGI-1 was near-solved. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,arc_agi_top,"ARC-AGI, top verified score (version changes: ARC-AGI-1 in 2024, ARC-AGI-2 from 2025)","% of tasks, semi-private set",2026,"84.6% (ARC-AGI-2, to date)",84.6,Feb 2026,ARC Prize Foundation verification of Gemini 3 Deep Think (upgrade),https://officechai.com/ai/gemini-3-deep-think-benchmarks-arc-agi,MEDIUM,"About $13.62 per task. BenchLM lists GPT-5.6 Sol at 92.5% (29 Jul 2026), but whether that run was ARC Prize-verified is unclear. Attention has moved to ARC-AGI-3, where frontier models score below 1% in the sources seen." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,mmlu_saturation,"MMLU, frontier score (saturation marker)",% correct,2024,92.3%,92.3,Sep 2024,OpenAI (o1 announcement),,MEDIUM,"Vendor figure, not re-verified in this session. MMLU launched in 2020 with frontier accuracy near 32%; its practical ceiling is about 95% because of label errors." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,mmlu_saturation,"MMLU, frontier score (saturation marker)",% correct,2025,>92% (saturated; labs report MMLU-Pro instead),,2025,Stanford HAI AI Index 2026,https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance,MEDIUM,No single top score tracked: model cards stopped headlining MMLU during 2025. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,mmlu_saturation,"MMLU, frontier score (saturation marker)",% correct,2026,>92% for every frontier system,,Q1 2026,"Stanford HAI AI Index 2026, as summarised on Report AI",https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance,MEDIUM,"Successor MMLU-Pro is itself near saturation: DeepSeek V4-Pro reports 87.5 (Apr 2026, page 964)." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,us_china_gap_arena,"Lead of the top US model over the best Chinese model, Chatbot Arena / LMArena",% score gap,2024,9.26%,9.26,Jan 2024,Stanford HAI AI Index 2025,https://hai.stanford.edu/ai-index/2025-ai-index-report/technical-performance,MEDIUM,"In 2023 the gap on major benchmarks was 17.5–31.6 percentage points (Stanford HAI, page 954)." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,us_china_gap_arena,"Lead of the top US model over the best Chinese model, Chatbot Arena / LMArena",% score gap,2025,1.70%,1.7,Feb 2025,Stanford HAI AI Index 2025,https://hai.stanford.edu/ai-index/2025-ai-index-report/technical-performance,MEDIUM,DeepSeek-R1 briefly matched the top US model in February 2025 (page 89). https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,us_china_gap_arena,"Lead of the top US model over the best Chinese model, Chatbot Arena / LMArena",% score gap,2026,2.7%,2.7,Mar 2026,Stanford HAI AI Index 2026,https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance,MEDIUM,"On a different instrument, Artificial Analysis's Intelligence Index, Alibaba's Qwen3.8-Max scored 58 against 58.9 for GPT-5.6 Sol on 3 Sep 2026 (page 1654), with Qwen3.8-Max's score moving 53→56→58 across re-runs." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,open_closed_gap_arena,"Lead of the top closed-weight model over the best open-weight model, Chatbot Arena / LMArena",% score gap,2024,8.04%,8.04,Jan 2024,Stanford HAI AI Index 2025,https://hai.stanford.edu/ai-index/2025-ai-index-report/technical-performance,MEDIUM,"Fell to 0.5% by Aug 2024 (Stanford HAI 2026, page 861)." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,open_closed_gap_arena,"Lead of the top closed-weight model over the best open-weight model, Chatbot Arena / LMArena",% score gap,2025,1.70%,1.7,Feb 2025,Stanford HAI AI Index 2025,https://hai.stanford.edu/ai-index/2025-ai-index-report/technical-performance,MEDIUM,Same 1.70% value as the US–China gap that month; the two are separate measurements. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,open_closed_gap_arena,"Lead of the top closed-weight model over the best open-weight model, Chatbot Arena / LMArena",% score gap,2026,3.3%,3.3,Mar 2026,Stanford HAI AI Index 2026,https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance,MEDIUM,"Gap widened again from its 2024–25 low. Anthropic, xAI, Google, OpenAI, Alibaba and DeepSeek sit within 25 Elo points of one another on Arena." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,aa_open_weight_gap,"Gap between best open-weight and best proprietary model, Artificial Analysis Intelligence Index",index points,2024,,,,,,,No comparable reading found; index versions changed. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,aa_open_weight_gap,"Gap between best open-weight and best proprietary model, Artificial Analysis Intelligence Index",index points,2025,13 points,13,~Apr 2025,Artificial Analysis (DeepSeek V3 0324 at 22 vs Claude 3.7 Sonnet at 35),https://artificialanalysis.ai/articles/recent-open-weights-model-launches,MEDIUM,Earlier index version; scores are not comparable with 2026 versions. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,aa_open_weight_gap,"Gap between best open-weight and best proprietary model, Artificial Analysis Intelligence Index",index points,2026,12 points (v5); 6 points on v4 in Sep,12,Oct 2026,"Artificial Analysis Intelligence Index v5, via secondary coverage",https://cybernews.com/ai-news/artificial-analysis-leaderboard-update-open-weight-ai-ranking-hit/,CONTESTED,"On v4.x in Sep 2026 the gap was about 6 points (Kimi K3 59.7 vs Claude Fable 5.1 65.7). Index v5 (Oct 2026, mostly private test sets) shows 12 points if Kimi K3.1 Pro counts as open, although its weights were not yet published; counting only published weights the best open model scores about 32 against 58 for Claude Opus 5.5, a ~26-point gap (DERIVED: 58 − 32). Artificial Analysis described the change as the gap going from 3 to 12 points. All 10 top open-weight models were Chinese in April 2026; Nvidia's Nemotron 3.5 Super VL led US open models at 26 in October." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,open_weight_lag_months,"Time lag of best open-weight model behind the closed frontier, Epoch Capabilities Index",months,2024,,,,,,,"Epoch publishes a multi-year average (Jan 2023–Oct 2025), not a 2024-only figure; see 2025." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,open_weight_lag_months,"Time lag of best open-weight model behind the closed frontier, Epoch Capabilities Index",months,2025,~3 months (average Jan 2023–Oct 2025),3,Oct 2025,Epoch AI,https://epoch.ai/data-insights/open-closed-eci-gap,MEDIUM,"Period average, not a single-year reading." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,open_weight_lag_months,"Time lag of best open-weight model behind the closed frontier, Epoch Capabilities Index",months,2026,"~4 months (average since Jan 2026, to date)",4,2026,Epoch AI,https://epoch.ai/data-insights/open-closed-eci-gap,MEDIUM,"Slight widening, consistent with Stanford HAI's 3.3% Arena gap." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,notable_models_us,Notable AI models produced by US-based institutions,models per year,2024,40,40,2024 (published Apr 2025),Stanford HAI AI Index 2025 (Epoch AI data),https://hai.stanford.edu/ai-index/2025-ai-index-report/research-and-development,MEDIUM,Europe produced 3 in 2024 (down from 21 in 2023); the US produced 61 in 2023. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,notable_models_us,Notable AI models produced by US-based institutions,models per year,2025,50 or 59,50,2025 (published Apr 2026),Stanford HAI AI Index 2026,https://hai.stanford.edu/ai-index/2026-ai-index-report/research-and-development,CONTESTED,"The Chapter 1 PDF says 50 (with South Korea third at 5); Stanford's R&D web page says 59. The difference may reflect a revised Epoch dataset. num uses the PDF figure. Either way the count is above 2024's 40, while Stanford says new releases declined across major regions on its own basis — check before citing." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,notable_models_us,Notable AI models produced by US-based institutions,models per year,2026,,,,,,,Not yet published; expected in the AI Index 2027 (spring 2027). https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,notable_models_china,Notable AI models produced by China-based institutions,models per year,2024,15,15,2024 (published Apr 2025),Stanford HAI AI Index 2025 (Epoch AI data),https://hai.stanford.edu/ai-index/2025-ai-index-report/research-and-development,MEDIUM,Also 15 in 2023. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,notable_models_china,Notable AI models produced by China-based institutions,models per year,2025,30 or 35,30,2025 (published Apr 2026),Stanford HAI AI Index 2026,https://hai.stanford.edu/ai-index/2026-ai-index-report/research-and-development,CONTESTED,PDF: 30; web page: 35. On either count China's share of US output rose from 38% (15/40) to 59–60% (30/50 or 35/59) — DERIVED. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,notable_models_china,Notable AI models produced by China-based institutions,models per year,2026,,,,,,,Not yet published. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,gpt35_class_inference_cost,Inference price of GPT-3.5-level performance (MMLU 64.8),$ per million tokens,2024,$0.07,0.07,Oct 2024,Stanford HAI AI Index 2025,https://hai.stanford.edu/ai-index/2025-ai-index-report/technical-performance,MEDIUM,"Down from $20.00 in Nov 2022: a more than 280-fold decline (DERIVED: 20 / 0.07 ≈ 286). Intermediate points on page 421 ($1.80 Aug 2023, $0.18 Jun 2024) are illustrative." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,gpt35_class_inference_cost,Inference price of GPT-3.5-level performance (MMLU 64.8),$ per million tokens,2025,,,,,,,Stanford HAI did not extend this specific series in the material found; see median_token_price for its 2025 measure. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,gpt35_class_inference_cost,Inference price of GPT-3.5-level performance (MMLU 64.8),$ per million tokens,2026,,,,,,,Not published. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,median_token_price,Median price per million tokens across AI models,$ per million tokens,2024,$2.10,2.1,2024,"Stanford HAI AI Index 2026, via secondary coverage",https://www.remio.ai/post/stanford-ai-index-shows-cheaper-models-fiercer-competition,MEDIUM,Single secondary source; not confirmed in the report text. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,median_token_price,Median price per million tokens across AI models,$ per million tokens,2025,$0.85,0.85,2025,"Stanford HAI AI Index 2026, via secondary coverage",https://www.remio.ai/post/stanford-ai-index-shows-cheaper-models-fiercer-competition,MEDIUM,"A 60% fall (DERIVED: 1 − 0.85/2.10). A median across listed models, not usage-weighted and not at fixed capability." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,median_token_price,Median price per million tokens across AI models,$ per million tokens,2026,,,,,,,Next reading due in AI Index 2027. For 2026 market pricing see silicon_data_token_index (different method). https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,silicon_data_token_index,Silicon Data LLM Token Expenditure Index (usage-weighted market price),$ per million tokens,2024,,,,,,,Index launched December 2025. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,silicon_data_token_index,Silicon Data LLM Token Expenditure Index (usage-weighted market price),$ per million tokens,2025,≈$1.0–1.1 (launch level),,Dec 2025,Report AI calculation from Bloomberg-reported movement,https://aiweekly.co/alerts/silicon-data-token-index-drops-20-from-may-bloomberg-warns,DERIVED,Silicon Data does not publish the launch value openly. Bloomberg reported the index had nearly doubled from its December inception to its late-May 2026 peak of $2.07: 2.07 / ~1.9–2.0 ≈ $1.04–1.09. Approximate. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,silicon_data_token_index,Silicon Data LLM Token Expenditure Index (usage-weighted market price),$ per million tokens,2026,$0.97 (to date; below $1 again in Oct),0.97,1 Sep 2026,"Silicon Data, via CNBC",https://www.cnbc.com/2026/09/01/ai-token-prices-lows.html,MEDIUM,"Axios (22 Oct 2026) said it fell below $1, 54% under the $2.07 late-May peak — implied ≈ $0.95 (DERIVED: 2.07 × 0.46). That a 'recent' fall below $1 followed a $0.97 reading on 1 Sep implies it rose above $1 in between, or Axios is lagging (flagged CONTESTED in research notes). Silicon Data's own segment readings on 22 Apr 2026: frontier $4.20, open-weight $0.85 per million tokens." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,price_decline_fixed_capability,Rate of price decline at a fixed capability level,× per year,2024,~10× per year,,Nov 2024,"a16z, LLMflation",https://a16z.com/llmflation-llm-inference-cost/,MEDIUM,Methods differ across the three years; do not chart as one series. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,price_decline_fixed_capability,Rate of price decline at a fixed capability level,× per year,2025,~50× per year median; ~200× per year since Jan 2024,,Mar 2025,"Epoch AI, LLM inference price trends",https://epoch.ai/data-insights/llm-inference-price-trends,MEDIUM,Range 9× to 900× per year depending on benchmark and threshold. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,price_decline_fixed_capability,Rate of price decline at a fixed capability level,× per year,2026,~47% per quarter (≈13× per year),,2026,"Epoch AI, via secondary coverage",https://pasqualepillitteri.it/en/news/21044/epoch-ai-ai-prices-fall-47-percent-quarter,MEDIUM,Single secondary source. Annualisation is DERIVED: 1 / (1 − 0.47)^4 ≈ 12.7×. Not a slowdown signal: different method and benchmark set from the 2025 figure. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,anthropic_top_tier_price,Anthropic Opus-tier list price (input / output),$ per million tokens (num = output),2024,$15 / $75 (Claude 3 Opus),75,Mar 2024,Anthropic pricing,https://www.anthropic.com/pricing,MEDIUM,Published list price at launch; not re-verified in this session. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,anthropic_top_tier_price,Anthropic Opus-tier list price (input / output),$ per million tokens (num = output),2025,$5 / $25 (Claude Opus 4.5),25,24 Nov 2025,Anthropic,https://www.anthropic.com/news/claude-opus-4-5,MEDIUM,"Two-thirds cut from the $15 / $75 of earlier Opus 4.x models. Opus 4.8 kept $5 / $25 (page 421, verified Jul 2026)." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,anthropic_top_tier_price,Anthropic Opus-tier list price (input / output),$ per million tokens (num = output),2026,$4 / $20 (Claude Opus 5.5),20,1 Oct 2026,"Anthropic pricing page, verified by Report AI",https://www.anthropic.com/pricing,HIGH,Verified at source by Report AI on 1 Oct 2026 (page 421). A new tier sits above Opus: Claude Fable 5.1 at $10 / $50. Batch halves prices; Opus 5.5 fast mode costs 2×. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,openai_flagship_launch_price,"OpenAI newest flagship, list price at launch (input / output)",$ per million tokens (num = output),2024,$15 / $60 (o1),60,Dec 2024,"OpenAI pricing, as cited on Report AI",https://openai.com/api/pricing/,MEDIUM,"Reasoning flagship. GPT-4o, the general model, listed at $2.50 / $10." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,openai_flagship_launch_price,"OpenAI newest flagship, list price at launch (input / output)",$ per million tokens (num = output),2025,$1.25 / $10 (GPT-5),10,7 Aug 2025,OpenAI pricing at GPT-5 launch,https://openai.com/api/pricing/,MEDIUM,Not re-verified in this session. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,openai_flagship_launch_price,"OpenAI newest flagship, list price at launch (input / output)",$ per million tokens (num = output),2026,$2 / $10 (GPT-6 Sol),10,22 Sep 2026,"OpenAI, via trade coverage",https://www.iclarified.com/102351/openai-cuts-api-prices-50-with-gpt-6-sol-and-luna,MEDIUM,"Launched with a 50% cut versus GPT-5.6 Sol, which listed at $5 / $30 in Jul 2026 (page 964; one price sheet shows $4 / $20). GPT-6 Astra (3 Sep 2026) is priced about 5× Sol. GPT-6.1 Terra (26 Oct 2026) lists at $1 / $5. OpenAI's pricing page was not checked." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,anthropic_small_tier_price,Anthropic Haiku-tier (smallest model) list price (input / output),$ per million tokens (num = output),2024,$0.80 / $4 (Claude 3.5 Haiku),4,Nov 2024,Anthropic pricing,https://www.anthropic.com/pricing,MEDIUM,Raised from Claude 3 Haiku's $0.25 / $1.25. Not re-verified in this session. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,anthropic_small_tier_price,Anthropic Haiku-tier (smallest model) list price (input / output),$ per million tokens (num = output),2025,$1 / $5 (Claude Haiku 4.5),5,Oct 2025,"Anthropic pricing, verified by Report AI",https://www.anthropic.com/pricing,HIGH,Listed as unchanged on page 421 (verified 1 Oct 2026). https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,anthropic_small_tier_price,Anthropic Haiku-tier (smallest model) list price (input / output),$ per million tokens (num = output),2026,$0.10 / $0.50 (Claude Haiku 5.5),0.5,7 Oct 2026,"Anthropic, via TechRepublic / Decrypt",https://www.techrepublic.com/article/news-anthropic-claude-haiku-5-5-api-pricing/,MEDIUM,"For prompts up to 100K tokens; $0.50 / $2.50 above. Anthropic says the same text uses ~30% more tokens with the new tokenizer, and whether the cut is 90% or a ~75% average is disputed. 1M-token context." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,deepseek_flagship_price,"DeepSeek flagship list price, open-weight price floor (input / output)",$ per million tokens (num = output),2024,$0.27 / $1.10 (DeepSeek-V3),1.1,Dec 2024,DeepSeek API pricing,https://api-docs.deepseek.com/,MEDIUM,Standard price after a launch promotion; not re-verified in this session. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,deepseek_flagship_price,"DeepSeek flagship list price, open-weight price floor (input / output)",$ per million tokens (num = output),2025,$0.55 / $2.19 (DeepSeek-R1),2.19,20 Jan 2025,DeepSeek,https://api-docs.deepseek.com/news/news250120,MEDIUM,"Reasoning model, about 96% cheaper than OpenAI o1's $15 / $60 (page 861). Mixes model types year to year: V3 (2024), R1 (2025), V4-Pro (2026)." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,deepseek_flagship_price,"DeepSeek flagship list price, open-weight price floor (input / output)",$ per million tokens (num = output),2026,~$0.87 output (DeepSeek V4-Pro),0.87,24 Apr 2026,"DeepSeek, as cited on Report AI",https://api-docs.deepseek.com/,MEDIUM,"MIT-licensed, 1.6T parameters. V4.1-Flash lists at $0.15 / $0.60 off-peak and $0.30 / $1.20 peak (Oct 2026)." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,largest_training_compute,Largest known training run (compute),FLOP,2024,~3.8×10^25 FLOP (Llama 3.1 405B); no model above 10^26,3.8e+25,Jul 2024,"Epoch AI, Large-Scale AI Models database",https://epoch.ai/data/large-scale-ai-models,MEDIUM,Largest 2024 model with a public estimate; closed 2024 models may be larger. Epoch says the first model estimated above 10^26 FLOP was Grok 3 (Feb 2025). https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,largest_training_compute,Largest known training run (compute),FLOP,2025,~5×10^26 FLOP (Grok 4),5e+26,Jul 2025,Epoch AI,https://epoch.ai/data-insights/grok-4-training-resources,MEDIUM,"Epoch's estimate, largely based on vague public statements from xAI. About 13× Llama 3.1 405B (DERIVED: 5e26 / 3.8e25)." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,largest_training_compute,Largest known training run (compute),FLOP,2026,,,,,,,No public compute estimates found for 2026 frontier models; public databases lack estimates for closed models after mid-2025. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,frontier_training_cost,Estimated cost to train a frontier model,$ (compute cost of final training run),2024,~$170M (Llama 3.1 405B),170000000,Jul 2024 model; published Apr 2025,Stanford HAI AI Index 2025 (Epoch AI estimate),https://hai.stanford.edu/ai-index/2025-ai-index-report/research-and-development,MEDIUM,Widely repeated; not confirmed against the report text this session. Other estimates for the same model range $30–60M under different scopes. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,frontier_training_cost,Estimated cost to train a frontier model,$ (compute cost of final training run),2025,~$490M (Grok 4),490000000,Jul 2025 model; published Sep 2025,Epoch AI,https://epoch.ai/data-insights/grok-4-training-resources,MEDIUM,"Median estimate: 246M H100-hours at $1.90–2.20 per hour; also ~310 GWh and ~150,000 t CO2e. Stanford HAI 2026 puts Grok 4 training emissions at 72,816 t CO2e, about half Epoch's figure (CONTESTED on emissions). Contrast: DeepSeek reported $294K for R1's reinforcement-learning stage (Nature, Sep 2025), on top of a $5.58M final V3 pre-training run — not a total model cost." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,frontier_training_cost,Estimated cost to train a frontier model,$ (compute cost of final training run),2026,,,,,,,"No credible public estimate found for any 2026 frontier model. Anthropic's S-1 (26 Oct 2026) discloses ~$245B of future compute and lease commitments, not per-model training cost." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,flagship_context_window,"Context window, median of OpenAI, Anthropic and Google flagships",tokens,2024,200K,200000,Dec 2024,"Vendor model documentation (GPT-4o 128K, Claude 3.5 Sonnet 200K, Gemini 1.5 Pro 2M)",,DERIVED,Median of the three flagships. Gemini 1.5 Pro's 2M window (Jun 2024) was the largest in a frontier closed model. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,flagship_context_window,"Context window, median of OpenAI, Anthropic and Google flagships",tokens,2025,400K,400000,Dec 2025,"Vendor model documentation (GPT-5 400K, Claude Opus 4.5 200K, Gemini 3 Pro 1M)",,DERIVED,"Median of the three. Largest advertised window in any notable release: Meta's open-weight Llama 4 Scout, 10M tokens (Apr 2025)." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,flagship_context_window,"Context window, median of OpenAI, Anthropic and Google flagships",tokens,2026,~1M (to date),1050000,Oct 2026,"Vendor documentation and trade coverage (GPT-6 Sol 1.05M, Claude Fable 5 / Haiku 5.5 1M, Gemini 3.1 Pro 2M)",,DERIVED,"Median of the three (GPT-6 Sol 1,050,000). Opus 5.5's window is 1M tokens (Anthropic models overview, checked 15 Nov 2026). Largest advertised by any model remains Llama 4 Scout's 10M tokens (Apr 2025); an earlier '5M-token Llama 5' figure was unsupported and corrected on 15 Nov 2026." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,imo_ai_result,Best AI result at the International Mathematical Olympiad,points out of 42,2024,28/42 (silver standard),28,Jul 2024,Google DeepMind (AlphaProof + AlphaGeometry 2),https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/,MEDIUM,Problems were formalised by humans; not a natural-language result. Not re-verified this session. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,imo_ai_result,Best AI result at the International Mathematical Olympiad,points out of 42,2025,"35/42 (gold, officially certified)",35,Jul 2025,"Google DeepMind (Gemini Deep Think), certified by the IMO",https://deepmind.google/discover/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/,MEDIUM,"5 of 6 problems, in natural language within contest time. OpenAI's experimental model also reported 35/42 but was self-graded, not IMO-certified." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,imo_ai_result,Best AI result at the International Mathematical Olympiad,points out of 42,2026,42/42 (perfect scores reported),42,Jul 2026,"Trade press (36Kr, NewsBytes, Xenospectrum)",https://www.digitalapplied.com/blog/imo-2026-perfect-scores-ai-benchmark-saturation,MEDIUM,"Several systems reported perfect scores at IMO 2026 in Shanghai, including Huawei's Celia, Xiaohongshu's dots-note-3.0, an xhigh GPT-5.6 Sol and Claude Fable 5 (single outlet for the last two). The six systems credited with 42/42 were not graded the same way, and no official IMO statement was found." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,energy_per_query,Energy per typical AI text query,watt-hours,2024,"~0.3 Wh (GPT-4o, estimate)",0.3,Feb 2025 estimate of a 2024 model,Epoch AI,https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use,CONTESTED,"The widely circulated 2023–24 estimate was ~2.9 Wh per query, about 10× higher; Epoch's re-estimate assumed newer hardware and shorter outputs. Methods differ across years." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,energy_per_query,Energy per typical AI text query,watt-hours,2025,0.24 Wh (median Gemini text prompt),0.24,Aug 2025 (measured May 2025),Google technical report,https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference,MEDIUM,Self-reported; also 0.26 mL water and 0.03 g CO2e. Sam Altman cited ~0.34 Wh per ChatGPT query (Jun 2025) without a method. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,energy_per_query,Energy per typical AI text query,watt-hours,2026,0.31 Wh median; ~3.9 Wh for long reasoning responses,0.31,2026,"Microsoft Research study in Joule, via secondary coverage",https://intuitionlabs.ai/articles/ai-inference-energy-joules-per-task,MEDIUM,"Single study; reasoning-style responses use roughly 13× the median. No new vendor per-query disclosure was found for 2026, and no regulator mandates one as of Sep 2026." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,frontier_releases,Notable frontier model releases,releases (list),2024,Claude 3 (Mar); GPT-4o (May); Claude 3.5 Sonnet (Jun); Llama 3.1 405B (Jul); o1-preview (Sep) and o1 (Dec); Gemini 2.0 Flash (Dec); DeepSeek-V3 (26 Dec),,2024,Vendor announcements,,MEDIUM,"Editorial selection, not a count. o3 was announced as a preview in Dec 2024 but released in Apr 2025." https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,frontier_releases,Notable frontier model releases,releases (list),2025,"DeepSeek-R1 (20 Jan); Claude 3.7 Sonnet, GPT-4.5 (Feb); Gemini 2.5 Pro (Mar); Llama 4, o3 (Apr); Claude 4 (May); Grok 4 (Jul); GPT-5 (7 Aug); Gemini 3 Pro (18 Nov); Claude Opus 4.5 (24 Nov); Mistral Large 3 (Dec)",,2025,Vendor announcements,,MEDIUM,Editorial selection. https://report-ai.org/indexes/state-of-ai/state-of-ai-models-prices/,2026-11-11,models,frontier_releases,Notable frontier model releases,releases (list),2026,Gemini 3.1 Pro (Feb); Meta Muse Spark (8 Apr; closed weights); DeepSeek V4-Pro (24 Apr); Gemini 3.5 Flash (19 May); Claude Opus 4.8 (28 May); Claude Fable 5 (9 Jun; suspended 12 Jun–1 Jul under export controls); Grok 4.5 (8 Jul); GPT-5.6 Sol (9 Jul); Qwen3.8-27B (14 Aug); Claude Fable 5.1 and Mythos 5.1 (1 Sep); GPT-6 Astra (3 Sep); GPT-6 Sol/Luna and Claude Opus 5.5 (22 Sep); Claude Haiku 5.5 (7 Oct); Gemini 4 Argon GA (13 Oct); Meta Muse Spark Open (15 Oct); Mistral Large 4 weights (29 Oct),,to 8 Nov 2026,Vendor announcements via Report AI pages and trade coverage,,MEDIUM,Dates for 2026 are from trade coverage; most vendor pages could not be opened from our environment. Kimi K3 (Moonshot) release date not confirmed.