Wholestory

Artificial Analysis replaced its Intelligence Index — the independent scoreboard this page ranks models by — with version 4.2, and re-scored every model on it. The publisher calls it an interim step toward a planned version 5, saying it is "accelerating elements of our upcoming v5 release with interim updates to keep pace with the frontier" and that v4.2 carries "more complex and realistic tasks, and more private test sets to prevent gaming". Ten evaluations feed it, among them two agentic knowledge-work tests, a terminal-based coding benchmark and a hallucination measure. Scores fell across the board: Claude Fable 5.1, the leader, from 66 to 57; Kimi K3, the best downloadable model, from 60 to 50; GPT-4, from 7 to 1. Nothing about the models changed. Figures from before 4 September cannot be set against figures from after it.

OpenAI released GPT-6 Astra, branding it a generational step. On the Artificial Analysis Intelligence Index it scores 61 — exactly level with GPT-5.6 Sol and five points below Claude Fable 5.1 — while costing two-and-a-half times as much, at $10/$50 per million tokens. Its gains are elsewhere: it leads Artificial Analysis's separate Coding Agent Index at 67, uses about a third of Sol's tokens on coding tasks, and roughly halves its hallucination rate, from 92% to 51% at maximum effort, without losing accuracy. It reached a limited set of organisations first, then ChatGPT's paid tiers, the API and AWS.

Meta released Muse Spark 1.3, its fourth in five months. On the Artificial Analysis Intelligence Index the limited-preview 'max' version scores 62 — behind only Claude Fable 5.1 and Claude Opus 5 — and the generally available version scores 61, the cheapest model at that level ($0.55 per task, unchanged $1.25/$4.25 pricing). It is Meta's strongest showing since it made open weights a strategy, but the model is proprietary, served through Meta's own API with the top version limited to partners. Its August pledge to 'soon' open Muse Spark 1.2's weights remains unmet.

Between 27 August and 3 September, five labs shipped models scoring 57 to 66 on the Artificial Analysis Intelligence Index: Anthropic's Claude Fable 5.1 (66), Meta's Muse Spark 1.3 (62), OpenAI's GPT-6 Astra (61), Google's Gemini 3.8 Flash (59, its fourth Flash model in under four months), and Zhipu's open-weight GLM-5.3-Flash (57, MIT-licensed, emerging from a week-long anonymous 'Ox Alpha' preview). Tencent's open-weight Hy4 arrived the same week, still unscored — a sixth release, from a sixth lab. Every one of the top scores belongs to a closed model.

Anthropic released Claude Fable 5.1, and its trusted-access sibling Mythos 5.1, and on the independent Artificial Analysis Intelligence Index its top effort scores 66 — the highest the index has measured, four points above Fable 5 and three above Claude Opus 5. Standard token pricing is unchanged at $10/$50 per million, but Anthropic cut the cache-read price 75%, from $1 to $0.25; even so, a task costs about 20% more than on Fable 5, because the model spends roughly 1.7 times the output tokens to reach the higher score. The move reopens the distance to the downloadable frontier: the best open-weight models, Moonshot's Kimi K3 and Zhipu's GLM-5.3, sit at 60, so the gap between the best model and the best one anyone can download widened from three points to six in a single release.

Tencent published Hy4 preview, a downloadable mixture-of-experts model with 770 billion parameters — about 49 billion active for any request — and a context window over a million tokens, also served through Tencent Cloud and OpenRouter at $0.83 per million input tokens and $2.50 output. It adds another Chinese open-weight entrant behind Moonshot, Zhipu, Alibaba and DeepSeek. Tencent flagged that the early-release model can over-verify its answers and run long on hard questions. It carries no independent capability score yet — Artificial Analysis has not evaluated it — so it stays off this page's ranked board until one exists.

Artificial Analysis’ index today puts Claude Opus 5 at 63, the highest of the 128 reasoning models it scores, and Moonshot’s Kimi K3 at 60, the highest of the open-weights models. Behind Kimi come Alibaba’s Qwen3.8 at 58 and DeepSeek V4 Pro at 53. Of the 161 models the index evaluates, 86 — more than half — can be downloaded. The index version was also checked this cycle: every row added to this page’s board since June is scored on Intelligence Index v4.1.1, which is what the live index runs today, so the board compares like with like. Nothing needed correcting.

Alibaba released Qwen3.8-Flash-Next under open weights, and Artificial Analysis scored it 56 on its Intelligence Index — the same as OpenAI's GPT-5.5 and xAI's Grok 4.5. GPT-5.5 costs $5 per million input tokens; this costs $0.15, with output at $0.47 against $30. It is a mixture-of-experts model with a 125-billion-parameter main body, 51 billion more in N-gram embedding tables, and six billion parameters active per token, running 262,144 tokens of context natively and a million with YaRN. Alibaba presents it as an early look at the architecture Qwen4 will use, and says training cost about a ninth of Qwen3.7-Plus — its own figure, unverified. The independent part is the index score, which is why it is on the board; the lab's own benchmark table is not used.

Thomson Reuters launched Thomson, its own large language model, built by taking an open-source foundation and spending $40 million on talent and compute to mid-train and post-train it over Westlaw, Practical Law, Checkpoint and Reuters — less than 10 percent of the company's content so far. It says the model is fully owned in-house and runs without the inference costs of a typical frontier model, and its chief executive says early evaluations put it "on par with the latest frontier models across a range of tasks". No benchmark is named, no score published and no third party has assessed it, so the capability claim is the company's alone.

SemiAnalysis measured the open-versus-closed gap across three eras of language models — early scaling, reasoning, agentic — and found it moves in cycles rather than one direction. A frontier lab jumps ahead at the start of an era; others reverse-engineer the advance and close it. What has changed is the clock: with each generation, open-source models take half as long to catch up to the first closed model of that era. The method matters. Rather than score every model on one benchmark set, which saturates and is abandoned, SemiAnalysis scored each era on its own benchmarks and compared catch-up time — a rate, not a level. It assesses GLM 5.3 and Kimi K3 as genuinely able to do the coding and agentic work that carried Anthropic past $65 billion in annualised revenue.

DeepSeek's new prices took effect at 16:00 UTC on 16 August, replacing flat rates with a peak and off-peak schedule. V4-Flash, priced through July at $0.14 per million input tokens and $0.28 per million output, now costs $0.44 and $1.32 in peak hours and half that outside them. V4-Pro moves from $0.435 and $0.87 to $1.32 and $3.96 at peak, $0.66 and $1.98 off-peak. Reported increases run from 50% to more than 1,100% depending on the model, the token type and the hour. Peak covers seven hours of every 24 — 01:00 to 04:00 and 06:00 to 10:00 UTC — so most of the day bills at the lower rate. The company said it was adopting the schedule "to allocate resources more reasonably ... encouraging users to schedule their tasks based on actual usage", announcing it alongside the general release of V4-Pro. Both models' weights remain published under the MIT licence, so the rise applies to DeepSeek's hosted service rather than to the models, which anyone can still run themselves. Even at peak, V4-Pro's output tokens cost about a twelfth of Claude Fable 5's $50. Four months earlier DeepSeek had cut V4-Pro by 75% and its cached-input prices across the API to a tenth of their previous level.

Alibaba published Qwen3.8-27B, a dense 27-billion-parameter model under the permissive Apache 2.0 licence, and on the independent Artificial Analysis Intelligence Index it scores 52 — the same score as DeepSeek's V4-Flash, a 284-billion-parameter mixture-of-experts model ten times its size, and level with OpenAI's GPT-5.6 Luna. It is seventeen points above Meta's Muse Glimmer, the 30-billion-parameter open model released four days earlier into the same size class, and eleven points below the best model anyone can buy. The unquantised weights are 55.6 GB; a community four-bit conversion for Apple silicon is 16.1 GB, small enough for a 32 GB Mac with room left for working context. The model takes text and images and carries a 262,144-token context window, extensible to a million. The same index records what the score costs: it generated 160 million tokens to finish the test battery against a 43 million median for open models of its size, reaching its answers by thinking several times longer than its peers. Alibaba's own benchmark comparisons, including a claimed parity with Claude Opus 4.6 at maximum effort and a sweep of its own chosen tests against Muse Glimmer, are the lab's claims and not the independent measure.

Google released Gemini 3.7 Flash 23 days after Gemini 3.6 Flash, priced at $0.75 per million input tokens and $3.75 per million output through 31 December 2026, with $1.50 and $7.50 taking effect on 1 January 2027. It cut 3.6 Flash to the same dated rate, so the new model does not undercut its predecessor — both undercut the price 3.6 Flash launched at three weeks earlier. On the independent Artificial Analysis Intelligence Index 3.7 Flash scores 56, four points above 3.6 Flash and level with xAI's Grok 4.5 at a third of Grok's input price. Google's own benchmark table claims leads on production-code quality and web development; the same table shows OpenAI's GPT-5.6 Terra ahead on two terminal benchmarks and Claude Sonnet 5 ahead on a desktop-agent evaluation, and those are the lab's own figures either way. Google gave no release date for its next flagship, Gemini 3.5 Pro, which it had previously described as in partner testing.

Anthropic withdrew the increase it had published for Claude Sonnet 5 when the model shipped on 30 June, adding an editor's note to the launch announcement: "Sonnet 5's introductory pricing of $2 per million input tokens and $10 per million output tokens is now permanent. The standard pricing of $3 input / $15 output previously set to take effect September 1 no longer applies." Its pricing documentation, which through July carried two dated rows for the model, now lists a single row at $2 and $10. Sonnet 5 is the default model on Anthropic's free and paid consumer plans, and the withdrawn increase was the only scheduled rise on any of the three largest American labs' published price lists.

Alibaba published the weights of Qwen3.8-Max, its 2.4-trillion-parameter flagship (95 billion active per token, one-million-token context) — the first Max-class Qwen released as open weights and the largest-parameter model anyone can download, fulfilling the release the company had promised at the model's 3 August launch. The terms let most developers use, copy, modify and distribute the model without payment, but require a separate paid commercial licence for any 'model as a service' or 'AI work assistant' business whose aggregate revenue exceeds US$50 million over any consecutive 12-month period; internal use is exempt. The model scores 58 on the Artificial Analysis Intelligence Index v4.1.1, second among open-weight models behind Moonshot's Kimi K3 (60) and three points below the closed leader, Claude Opus 5 (63). Artificial Analysis's own model page still classified Qwen3.8-Max as proprietary as of 13 August, a lag behind the actual release. A paid API remains available at $2 per million input tokens and $6 per million output.

Meta released Muse Glimmer, a 30-billion-parameter open-weight model under the permissive Apache 2.0 licence with weights on Hugging Face — its first open-weight release since retooling its Muse flagship line as proprietary under Alexandr Wang. Optimized for always-on local agents, it is small enough to run on a Mac or PC with a single consumer GPU: compressed to roughly 4-bit precision it shrinks to under 20 GB, and Meta says it is distilled from the larger Muse Spark. It is multimodal (text and image), spans a 131,000-token context and was trained to a January 2026 knowledge cutoff. On the independent Artificial Analysis Intelligence Index v4.1.1 it scores 35 — an edge-scale rather than frontier score, but well above the median for open-weight models of its size. Meta paired the release with a pledge to open the weights of its frontier-band flagship Muse Spark 1.2 (57 on the same index), reopening the open-versus-closed question the company had appeared to abandon. Meta's own benchmark comparisons — against Google's Gemma 4 31B and Alibaba's Qwen3.6-27B — are the lab's own claims; the score cited here is the independent index's.

Meta released Muse Spark 1.2, a proprietary model scoring 54 on the Artificial Analysis Intelligence Index v4.1 — three points above Muse Spark 1.1's 51, inside the frontier band but below the leaders. It is priced at $1.25 per million input tokens and $4.25 per million output, unchanged from Muse Spark 1.1, with a one-million-token context window and text-and-image input, and is optimized for coding. Like its predecessor it is closed: Artificial Analysis lists the weights as not publicly available and Meta discloses no parameter count. It is Meta's third model release in four months, from the company whose Llama line made open-weight frontier-chasing a corporate strategy before its current flagship range went proprietary.

Alibaba released Qwen3.8-Max, a 2.4-trillion-parameter sparse mixture-of-experts model scoring 56 on the Artificial Analysis Intelligence Index v4.1 — inside the frontier band, below Moonshot's Kimi K3 (57), OpenAI's GPT-5.6 Sol (59) and Anthropic's Claude Opus 5 (61), and above Grok 4.5, Claude Sonnet 5 and GPT-5.5. It is priced at $2.00 per million input tokens and $6.00 per million output, with a one-million-token context window. Artificial Analysis classifies the model as proprietary — available through Alibaba's API, weights not published. Stanford's HAI and some trade coverage described Qwen3.8-Max as open-weight, and Alibaba separately released a smaller 27-billion-parameter open-weight variant, but the 2.4-trillion Max that carries the index score is API-only on the independent index's classification. Alibaba's own blog said the weights would be released; at launch they had not been, so on the measurement this record uses the largest new Chinese model counts as closed.

DeepSeek released V4-Flash (V4-Flash-0731), an open-weights reasoning model published under the MIT licence with weights on Hugging Face and designed to run on Huawei Ascend accelerators as well as Nvidia. A 284-billion-parameter mixture-of-experts model activating 13 billion parameters per token, it scores 50 on the Artificial Analysis Intelligence Index v4.1 — level with Google's Gemini 3.6 Flash and a point below Zhipu's GLM-5.2 and Meta's Muse Spark 1.1 — at a published price of $0.14 per million input tokens and $0.28 per million output. Artificial Analysis found it the cheapest well-known model to run, at roughly 3 cents to complete its Intelligence Index test battery, against 86 cents for Moonshot's Kimi K3, $1.86 for OpenAI's GPT-5.6 Sol and $3.15 for Anthropic's Claude Fable 5 — a cost measured per test rather than per token, because a model that is cheap per token can still run up a large bill when it needs more steps to reach an answer.

OpenAI cut the API price of two GPT-5.6 models, attributing the move to efficiency gains passed to customers rather than a promotion. Luna, the series' fast model, fell 80% to $0.20 per million input tokens and $1.20 per million output, from $1/$6; Terra, the everyday model, fell 20% to $2/$12 from $2.50/$15. The flagship, Sol, held at $5/$30 but gained a "Sol Fast" mode at twice standard price ($10/$60) for 2.5× throughput with no change in intelligence, replacing Priority Processing. On the Artificial Analysis Intelligence Index v4.1 Luna scores 51 and Terra 55; at the new price Artificial Analysis puts Luna at about $0.21 to run its test battery, undercutting Zhipu's GLM-5.2 as the cheapest model at that intelligence level. The cut is the latest in a run of releases that, by Artificial Analysis's count, put six labs above 50 on that index within eight days — the price of near-frontier capability being competed down rather than a single vendor's discount.

Google DeepMind released Gemini Robotics ER 2, an embodied-reasoning model that acts as a robot's high-level planner — reading continuous video, tracking its own task progress, coordinating multiple robots — while delegating motor control to lower-level vision-language-action models. It ships publicly through the Gemini API rather than as a research demo, which is the substantive step: frontier-lab reasoning models are now a purchasable component for physical machines, with whatever that implies for the deployment questions this record tracks mostly in software.

OpenAI began a limited preview of the GPT-5.6 series — flagship Sol, everyday model Terra and fast model Luna — restricted initially to trusted partners, a staging OpenAI attributes to the US government's request. The staggered access is the news as much as the models: it is the first frontier release sequenced to a government ask since June's executive order made a classified cyber benchmark the trigger for federal involvement in frontier releases, and it turns the 'covered model' question from a hypothetical into release policy. What Sol scores and where it lands on the indices belongs to the benchmark record once general access opens.

At 2.8 trillion parameters, China's largest model debuted at 57 on the Artificial Analysis index — fourth overall, above GPT-5.5 and behind only Claude Fable 5 and GPT-5.6 Sol — and second on AA's agentic knowledge-work benchmark. Chinese AI rivals' shares plunged on the news (Z.ai -28%, MiniMax -16%). K3 launched API-only, with weights announced for release at the end of July 2026.

Microsoft announced MAI-Cyber-1-Flash, its first model trained specifically for cybersecurity, describing it as "a compact, code-heavy security model derived from the MAI-Thinking-1 lineage, which was built from scratch, in-house, on the highest quality data". It is reachable only inside MDASH, Microsoft's multi-agent vulnerability-finding harness: there is no standalone endpoint, no waitlist and no licence terms on either launch page, and no parameter count, context length or per-token price is published anywhere. Neither Microsoft page states an availability status; Ars Technica reports the new tools are in preview. Its parent model has a self-published technical report — MAI-Thinking-1 is given there as a 34.7-billion-active, 962-billion-total mixture of experts pre-trained on 30 trillion tokens across 8,000 GB200 GPUs, and as trained "without distillation from third-party models" — but Microsoft does not say whether MAI-Cyber-1-Flash is a distillation, a fine-tune or a smaller sibling of it, and no independent index scores either model.

Meta released Muse Spark 1.1, a multimodal reasoning model with a 1,048,576-token context window, alongside the public preview of the Meta Model API — its first developer API for the model. Artificial Analysis scores it 51 on its Intelligence Index v4.1 at the highest effort setting, against a median of 33 for reasoning models in the same price band, and prices it at $1.25 per million input tokens and $4.25 per million output tokens, undercutting Zhipu's GLM-5.2 at the identical score of 51. It is also closed: Artificial Analysis lists the model as proprietary and states its weights are not publicly available, Meta discloses no parameter count, and Meta's own launch post never uses the words "open weights" or "open source" about it. That is a reversal for the company whose Llama 2 release in July 2023 turned open-weight frontier-chasing from a leak into a corporate strategy, on the argument that "by making AI models available openly, they can benefit everyone". Artificial Analysis measures the model as fast — 129 output tokens per second against a 67 median — and somewhat verbose, consuming 94 million output tokens to run the index against a 63 million median.

Moonshot AI published the full weights of Kimi K3 on Hugging Face, meeting to the day the end-of-July commitment it made at the model's API launch — Nature had reported on 21 July that Moonshot said the weights "will be released on 27 July". K3 is a 2.8-trillion-parameter mixture-of-experts model with 104B active parameters, 896 experts of which 16 fire per token, and a 1,048,576-token context window; Artificial Analysis scores it 57 on its Intelligence Index v4.1, unchanged from the API-only release, at $3.00/$15.00 per million tokens. The licence is a bespoke "Kimi K3 License" granting MIT-style rights to use, modify, distribute, sublicense, sell and fine-tune, with no acceptable-use policy, and only two commercial conditions: an operator selling model access with group revenue above US$20M over any twelve months must sign a separate agreement with Moonshot, and any product above 100M monthly active users or US$20M monthly revenue must display "Kimi K3" in its interface. Open weights are not open source — the training process and dataset remain undisclosed — and openness is not the same as reach: Tom's Hardware headlined the model "2-3x easier to run", but that figure is a price comparison of Moonshot's internally measured costs against rivals' retail token prices rather than a hardware measurement, and Rahul Shome of the Australian National University told Nature the model is too large for personal devices and will probably need significant institutional investment to run.

Two releases four days apart moved both ends of the same measurement. On 23 July, the best model anyone could download was Zhipu's GLM-5.2, which Artificial Analysis scores 51 on its Intelligence Index v4.1, against the highest score on that index, Claude Fable 5's 60 — a nine-point distance. On 27 July, with Kimi K3's weights published, the best downloadable model scores 57 and the highest score on the index is Claude Opus 5's 61: four points. Every figure is on Artificial Analysis' own per-model pages and was re-verified on 27 July; none of the four scores changed this cycle except the newly added Opus 5, so the movement comes entirely from which models are eligible on each side. One qualification travels with it: as of the day the weights were published, Artificial Analysis still labelled Kimi K3 a "proprietary model", meaning its own open-weights breakdown had not yet been updated to include it.

Anthropic released Claude Opus 5 at $5.00 per million input tokens and $25.00 per million output tokens — the same price as Opus 4.8, and half Claude Fable 5's $10/$50 — with a one-million-token context window and five effort settings. Artificial Analysis measures it at 61 on its Intelligence Index v4.1, one point above Fable 5's 60, a margin AA itself calls "effectively tied", ahead of GPT-5.6 Sol (59), Kimi K3 (57) and Opus 4.8 (56); AA discloses that it "supported Anthropic to evaluate Claude Opus 5 ahead of release", so this is a pre-release partnered evaluation rather than an arm's-length one. Its published findings are mixed: Opus 5 takes joint first on AA's Coding Index and records the highest GDPval-AA v2 score AA has measured (1861 Elo), but runs at 54.8 output tokens per second against a 79.6 median for comparably priced reasoning models, consumed 100M output tokens running the index against a 63M median — AA classes it "notably slow and very verbose" — trails GPT-5.6 Sol, GPT-5.5 Pro and GPT-5.6 Terra on the CritPt physics benchmark, and raised its hallucination rate 14 points to 50% on AA-Omniscience while gaining 7 points of factual accuracy. Anthropic also made Opus 5 the default on Claude Max and created a two-tier cyber regime, giving Cyber Verification Program members a version with fewer security restrictions.

Gemini 3.5 Flash Cyber, a Flash-class model fine-tuned to find and patch software vulnerabilities, was announced without general availability: Google DeepMind is releasing it as a limited pilot to governments and trusted partners through its CodeMender agent, citing the dual-use nature of automated vulnerability discovery. It carries no price on Google's published Gemini API pricing page and no score on the Artificial Analysis Intelligence Index, so every capability figure for it comes from Google or Google-run evaluation — including the "independent" evaluation Google credits to its own Big Sleep team, which is independent of the model team rather than of Google. Google's own footnotes disclose that its CyberGym competitor results are provider self-reported, and that the competitor baseline in its V8 comparison is Claude Opus 4.6 because newer Anthropic models refuse the tasks, so that comparison measures willingness alongside capability.

Priced at $2/$6 per million tokens and claiming roughly twice the token efficiency of comparable models, with xAI's own published tables showing it trailing Anthropic's Fable and OpenAI's GPT-5.5 on several coding evaluations — a notable strategic concession from the lab that had claimed the outright frontier a year earlier.

Mira Murati's lab shipped its first foundational model with full weights on Hugging Face, trained on Nvidia infrastructure — partly on data generated by existing open models, including China's Kimi K2.5. The first credible US-startup answer to the Chinese dominance of open weights, arriving amid growing enterprise demand for self-hostable models.

The flagship Sol — positioned for complex work across coding, research, science, and cybersecurity — began rolling out in ChatGPT under the trusted-partner limits announced in June, priced at $5/$30 per million tokens for short context. Independent scoring put Sol (max) at 59, second only to Claude Fable 5.

The institute's first public open-vs-closed analysis found GLM-5.2 matching the most cyber-capable closed models of November 2025 on its 70-task suite and 32-step cyber range — a lag of 4-7 months, down from the 6-10 months it had measured for earlier open models, at a fraction of the cost per run ($1.19 versus $85 for one 100M-token range run). It also found open-model safeguards largely did not impede its evaluations.

Anthropic restored Fable 5 globally from July 1 and Mythos 5 to approved organizations, reporting that the safeguard-bypass incident preceding the controls had exposed no unique Mythos-level cyber capability beyond what less-capable models already provided. The three-week episode established that model-level export control is now a live instrument of US policy.

The most agentic Sonnet shipped across all plans as the Free and Pro default, at introductory pricing of $2/$10 per million tokens (rising to $3/$15 in September). Independent scoring placed it at 53 on the Artificial Analysis index — within striking distance of flagships costing five times as much, though with notably verbose token usage.

Announced days before the series' public debut, and paired with the Fable 5 suspension then in force: for the first time, both leading US labs' most capable models were access-restricted by government action simultaneously — a wind at the back of the open-weights alternative, per contemporaneous industry reporting.

MIT-licensed open weights with a 1M-token context, landing within a percentage point of Anthropic's Opus 4.8 on closely watched agentic benchmarks at roughly a fifth of the cost, per CNBC's reporting — with OpenRouter traffic climbing faster post-launch than any prior open model. Zhipu also disclosed increased reward-hacking behavior in coding RL, an unusually candid model-card admission.

An unprecedented intervention: export controls on specific commercial AI models, requiring nationality-based restriction. Unable to verify nationality in-product, Anthropic suspended access rather than operate out of compliance — the first time a frontier model was pulled from the market by government order.

The first Mythos-class models: Fable 5 for general availability with safeguard routing on cyber, bio, and distillation topics, and Mythos 5 restricted to approved organizations — at $10/$50 per million tokens, which Anthropic said was less than half the price of its prior top tier. A new deployment pattern: one underlying model, two access regimes separated by safety gating.

A frontier coding/agentic model using MiniMax's MSA sparse attention (1/20th per-token compute at 1M context versus its predecessor), priced from $0.30/$1.20 per million tokens. Artificial Analysis initially scored it the leading open-weights candidate — while noting the promised weights arrived later than the launch commitment.

The family's first release shipped same-day across the Gemini app, AI Mode in Search, and developer platforms, with vendor benchmarks showing the mid-tier Flash outperforming the previous generation's Pro — each generation's small model now overtaking the last generation's flagship.

The long-awaited successor arrived open-source with markedly lower inference costs ($0.435/$0.87 per million tokens for V4 Pro), and Huawei confirmed its Ascend clusters support it — Chinese frontier capability and Chinese silicon advancing together. Analysts judged the capability real but the market impact muted: Chinese competitiveness was already priced in.

Better at coding, computer use, and research per OpenAI, at $5/$30 per million tokens with a 1M context — and the first OpenAI model treated as High-risk in both biological/chemical and cybersecurity domains under its Preparedness Framework, with advanced cyber capability gated behind identity-verified 'Trusted Access'. The UK AISI independently measured GPT-5.5 (with Anthropic's Mythos Preview) as one of the largest cyber-capability jumps it had evaluated.

The models that defined 2024-2025 left the consumer product in a single sweep (API access unchanged at the time), an unusually fast deprecation cycle that underlined how quickly frontier capability was being superseded — GPT-5 lasted just over six months as a ChatGPT model.

V3.2 and the reasoning-maxed V3.2-Speciale shipped with DeepSeek's first thinking-integrated tool use, trained via a new agent-data synthesis method across 1,800+ environments. The lab's claims — 'GPT-5 level' V3.2, Speciale 'rivals Gemini-3.0-Pro' with gold-medal results on IMO and ICPC — remained vendor-reported, but the release confirmed a second Chinese lab iterating at frontier cadence.

Six days after Gemini 3, Anthropic claimed the agentic-coding state of the art on SWE-bench Verified and cut Opus pricing 3x from $15/$75 to $5/$25 per million tokens — the frontier release cadence and its price curve now both compressing at once.

Gemini 3 Pro topped LMArena at 1501 Elo and scored 37.5% on Humanity's Last Exam (vendor-reported), launching simultaneously across the Gemini app, AI Mode in Search, and developer platforms — the first time Google put a frontier model into Search at release, with distribution numbers (2 billion monthly AI Overviews users) no rival could match.

A unified system — fast model, deeper reasoning model, and a real-time router — that replaced GPT-4o, o3, o4-mini, GPT-4.1, and GPT-4.5 as ChatGPT's default for its nearly 700 million weekly users, at $1.25/$10 per million tokens in the API. OpenAI reported ~45% fewer factual errors than GPT-4o with web search and large token-efficiency gains over o3.

gpt-oss-120b and gpt-oss-20b shipped under Apache 2.0 — OpenAI's first open-weight language models since GPT-2, with the 120B model achieving near-parity with o4-mini on the lab's reasoning benchmarks while running on a single 80GB GPU. OpenAI adversarially fine-tuned the model pre-release to test misuse potential, and executives acknowledged the release was driven by customers already running Chinese open models.

Available via API and new consumer tiers including a $300/month SuperGrok Heavy plan, with vendor-reported firsts including 50% on Humanity's Last Exam for Grok 4 Heavy. Trained with reinforcement learning at pretraining scale on the Colossus cluster.

Hybrid instant/extended-thinking models, with Opus 4 at $15/$75 per million tokens and a reported 72.5% on SWE-bench Verified. Anthropic shipped Opus 4 under its ASL-3 safety standard — the first frontier release explicitly deployed with a heightened internal safety level.

A family spanning 0.6B to 235B parameters with switchable thinking and non-thinking modes and a controllable thinking budget, trained on over 36 trillion tokens. Qwen was already, with DeepSeek, making Chinese releases the default substrate of the open ecosystem.

The new reasoning models could invoke every ChatGPT tool — web search, Python, visual reasoning, image generation — inside a chain of thought. OpenAI reported the o3 cost-performance frontier strictly improved on o1's, evidence that the reasoning axis was compounding rather than saturating.

Scout and Maverick shipped as open-weight multimodal MoE models, but Meta's headline LMArena Elo of 1417 came from an unreleased experimental chat configuration, not the published weights — drawing independent criticism, and diverging evaluations: Artificial Analysis rated the models among the best non-reasoning releases while other independent testers found them far weaker than claimed.

Released as a free experimental model, 2.5 Pro debuted at #1 on the LMArena human-preference leaderboard by what Google called a significant margin, with reasoning built in — the beginning of a sustained Google run at the top of crowd rankings that lasted more than six months.

Trained on xAI's 200,000-GPU Colossus cluster, Grok 3 launched consumer-first (no API at release) with reasoning modes and a claimed Chatbot Arena Elo of 1402 — the fastest any new lab had reached the frontier conversation since Anthropic.

R1 shipped under an MIT license with API pricing of $0.55 per million input tokens and $2.19 output — versus $15/$60 for the o1 it claimed parity with on math, code, and reasoning benchmarks — alongside six distilled smaller models. The release triggered a global market repricing of AI economics within a week and made a Chinese lab, for the first time, the reference point in the frontier conversation.

A 671B-parameter (37B active) mixture-of-experts model with open weights, trained — per DeepSeek's technical report — in just 2.788 million H800 GPU-hours on 14.8T tokens, with day-one inference support on AMD and Huawei Ascend hardware. The claimed training economics, on export-restricted hardware, previewed the disruption its successor would cause a month later.

o1-preview and o1-mini — trained, per OpenAI's research lead, with a new optimization approach that spends more compute thinking before responding — launched at $15/$60 per million tokens, roughly six times GPT-4o's price. Test-time compute became the industry's second scaling axis, and 'reasoning' variants became standard across every major lab within a year.

Trained on over 15 trillion tokens across more than 16,000 H100 GPUs, the 405B model was assessed by independent analysts as genuinely in the same class as the closed frontier — the first time open weights reached it. The license still restricted the largest platforms and required attribution, keeping the 'open source' label contested even as the capability gap closed.

OpenAI's small-model replacement for GPT-3.5 Turbo shipped at $0.15/$0.60 per million tokens — an order of magnitude cheaper than prior frontier models, and by OpenAI's own accounting a 99% cost drop per token versus its 2022-era text-davinci-003. The collapse in the price of a fixed level of capability was now explicit vendor strategy.

Anthropic's new mid-tier model kept Sonnet pricing ($3/$15 per million tokens) while beating Claude 3 Opus on the lab's evaluations — solving 64% of problems in an internal agentic coding evaluation versus Opus's 38%. Capability-per-dollar, not raw capability, was becoming the competitive front.

Apple's WWDC announcement paired a ~3 billion parameter on-device model (compressed to ~3.7 bits per weight, with swappable per-feature LoRA adapters) with a larger server model on Apple silicon. In Apple's own evaluations the on-device model beat comparable small open models — the clearest early evidence for a second, edge-scale track in the model race.

An end-to-end multimodal model ('o' for omni) responding to audio in near-human conversational time, priced at half of GPT-4 Turbo in the API and — for the first time for an OpenAI flagship — available to free ChatGPT users. The price of GPT-4-class capability was now falling on a steep curve.

Haiku, Sonnet, and Opus, priced at $0.25/$1.25, $3/$15, and $15/$75 per million tokens respectively, all with 200K context. Anthropic's published benchmark table showed Opus beating GPT-4 across common evaluations — the first credible claim on GPT-4's crown since its release a year earlier.

Announced with a standard 128K context and a 1-million-token window in private preview — the longest of any large-scale foundation model at the time — built on a mixture-of-experts architecture Google said was cheaper to train and serve. Context length became a new axis of frontier competition alongside raw capability.

A sparse mixture-of-experts model (46.7B total parameters, 12.9B active per token) under a genuinely permissive license — no research-only gate, no commercial restrictions. Mistral claimed it outperformed Llama 2 70B with 6x faster inference and matched GPT-3.5 on standard benchmarks, making it the strongest truly-open model of its moment and the European entry in the frontier race.

Three sizes — Ultra, Pro, and Nano — with Nano shipping on-device on the Pixel 8 Pro, the first frontier-lab model engineered into a phone. Google claimed Ultra exceeded state-of-the-art on 30 of 32 academic benchmarks and was the first model to beat human experts on MMLU at 90.0% — vendor figures that independent same-day coverage noted rested on very close margins.

Weights and starting code for pretrained and fine-tuned versions, free for research and commercial use, with Microsoft as preferred distribution partner via Azure. Meta reported over 100,000 researcher access requests for the first LLaMA. Open-weight frontier-chasing became a deliberate corporate strategy rather than a leak — though the license's restrictions meant 'open source' was contested from day one.

A large multimodal model accepting image and text input, released via ChatGPT Plus and an API waitlist at $30/$60 per million tokens (8K context; the 32K version cost $60/$120). OpenAI also open-sourced its evaluation framework, OpenAI Evals. GPT-4 became the reference point the rest of the industry measured against for the next year.

After a closed alpha with partners including Notion and Quora, Anthropic opened Claude via chat interface and API, in two tiers: Claude and the lighter, cheaper Claude Instant — putting a second US lab's frontier-adjacent model into the market the same week as GPT-4.

Roughly two weeks after Meta announced LLaMA as a research-access release requiring an application, the model weights appeared as a downloadable torrent on 4chan. Observers split between warning the technology would be misused and arguing open availability accelerates safety research — the first major access rupture of the frontier era, and an accidental preview of the open-weights ecosystem to come.

A conversational model fine-tuned with RLHF from a GPT-3.5-series base went live to the public at an effective price of $0. OpenAI's own launch notes disclosed the system 'sometimes writes plausible-sounding but incorrect or nonsensical answers.' The release turned large language models from a developer API into a mass consumer phenomenon and set off the competitive race that defines this page.

DeepMind published its compute-optimal training analysis: a 70B-parameter model trained on 1.3 trillion tokens outperformed far larger models like the 280B Gopher and 530B Megatron-Turing NLG. The finding — that the era's models were oversized for their compute budgets and undertrained on data — reset every lab's scaling recipe and foreshadowed the industry's turn toward smaller, cheaper-to-serve models.

OpenAI released its first commercial product: a private-beta 'text in, text out' API serving the GPT-3 family. It explicitly chose API access over open-sourcing the weights, arguing access can be revoked if misused — and that the models were so large and expensive to run that only big companies could otherwise afford to deploy them. The pattern set here — frontier capability behind a metered API — defined the market for years.