Frontier launches, capabilities, and what they actually mean.
Inception's new model offers a 1M-token context and launch pricing of $0.10/$0.35 per million tokens. In Inception's own 12-benchmark comparison, Mercury 3 is highest on eight, including both agentic-search tests; GPT-6 Luna and Claude Haiku 5.5 lead on two each, and take both long-context tests between them.
MiniMax lists $0.15 per million input tokens and says inference is nearly three times faster. In its own 15-row benchmark table the model is top or joint top on three rows.
The start-up says its first model, a research preview, matches or beats larger models on some visual benchmarks; its own table shows it behind on others, and its launch pages do not say which countries can use the API.
Nvidia calls Super VL "the fastest open agentic model in its class — 5x faster than the next best"; by our count, its own model-card table shows Super VL with the top score on 1 of 42 scorable rows, and higher than Qwen3.5-122B-A10B on 27 of 42. The weights are under the OpenMDW-1.1 licence.
Meta calls it "the most capable US open-weight model to date" and says two M5 Ultra Mac Studios can hold it, without stating the precision. By our count, a rival scores higher on 13 of 20 capability rows in Meta's longer model-card table; in Meta's 11-row blog scorecard Muse Spark Open leads on six.
The base model is still being post-trained, and the post offers no access route yet. Microsoft says it beats four outside models on its own held-out tests; three of those were tested as post-trained models.
Google says its frontier model is now in the Gemini app for AI Pro and Ultra subscribers, 13 days after a launch that went first to vetted cyber defenders. A new “tasks” toggle comes with it, but Google’s help page says that, for now, tasks are not available in the UK or the European Economic Area.
Meta says Muse Wonder and Muse Wonder Full deliver "state-of-the-art performance on personal and professional agentic tasks"; in its own comparison table, rival models score higher on several rows. Muse Wonder costs $4 per million input tokens on Meta's API; Wonder Full is limited to Muse Code subscribers, and a full safety report is "available soon".
Anthropic says Haiku 5.5, released on 7 October, beats Haiku 4.5 and OpenAI's GPT-6 Luna on its own benchmark table while trailing Sonnet 5.5 on every row. Early independent tests broadly back the gains, flag heavy token use, and in one test find GPT-6 Luna cheaper.
Mistral says its largest model to date leads open models on some legal and cyber tests. Its own charts also show it behind GLM-5.3 and Kimi K3 on several coding and agent benchmarks, and behind Claude Opus 5 in a human coding rating, though narrowly ahead of it on a science-coding test.
The German company published the weights on Saturday 3 October under Apache 2.0, which its model card says covers only the weights and configuration files. On Aleph Alpha's own table, Kolibri leads overall among the mixture-of-experts models it compared; by our count, a rival scores higher on 37 of the 51 individual benchmarks.
A preprint from Pangram Labs and University of Maryland researchers says 31.1% of filtered August 2026 web tokens are labelled as AI-generated by Pangram's detector, and proposes a scaling law in which an AI token's value can turn negative.
In a vendor-run study, Tavus says 48% of 54 participants believed its new full-duplex video model was human, against 1 of 41 for its earlier system. The model is not yet available to customers.
Cohere released Embed 5 on 30 September with two tiers that, it says, share one embedding space. Every score in the launch post was reported by Cohere, and several, including its headline ViDoRe V3 lead, use RCP-nDCG@10, an evaluation method Cohere introduced the same day.
Google DeepMind published SynthID Bio on 30 September 2026: watermarks embedded in AI-designed protein sequences and AI-predicted structures, with the sequence method validated in the lab on three binding targets. The paper calls it a proof-of-concept, and is explicit about what the watermark cannot do.
ElevenLabs says Eleven v4 is its "most emotive" text-to-speech model yet, with more than 90 languages and a Turbo variant it puts at a median inference latency of ~100ms.
H Company has released two open-weight Holo4 models under different licences. Its own benchmark table puts Holo4 27B at 61.7% on OSWorld 2.0 against a public 81.8% for Anthropic's Opus 5.5, which H did not run itself, and says its model costs much less per task.
The MIT-licensed mixture-of-experts model drops full attention for a 1M-token context. The company's own benchmarks put it first in six of 13 panels, three of them two-row comparisons, and between 4th and 7th on the three coding panels with the largest fields.
Reuters, citing The Information, says a researcher found a flaw that could have exposed a Muse user's dedicated virtual machine, and that Meta is making a safety warning clearer.
Perceptron says its new model adds audio input, object tracking through video and tool use, and runs up to 4.7 times faster than its predecessor. The benchmark figures are Perceptron's own, and in one of its tables the model trails at least one rival on each of the four standard video tests.
The Science paper says RADAR, trained on more than 400,000 CT exams, raised 26 radiologists' diagnostic sensitivity by about 10% in a reader study. The code is open, but the weights are under a non-commercial licence.
Google's new voice models can reason while speaking, run tools in the background and switch between 97 languages mid-conversation. The benchmark claims are strong. The model card adds caveats: a January 2025 knowledge cutoff, and a frontier-safety assessment that found no meaningful new capabilities compared with Gemini 3.7 Flash.
The maker of Devin says its new SWE-2 gets within a single point of Anthropic's Fable 5.1 on its own coding benchmark, at about 64% lower cost — and it's built on top of a Chinese open model, Moonshot's Kimi K3.
V4.1 Flash launched today as the smallest model in DeepSeek's new architecture. From 14 September, requests to the previous V4 Pro flagship get routed to it — at the cheaper Flash price.
MAI-Transcribe-2, released this week by Microsoft's own model division, tops the main multilingual benchmark, runs several times faster than its rivals, and lists at $0.10 an hour of audio. It is a direct shot at the transcription businesses of OpenAI, Google and ElevenLabs — and another sign of Microsoft quietly building models to lean less on its partner, OpenAI.
The Institute of Foundation Models, MBZUAI's research arm, has released K2 Horizon: six Apache-2.0 models from 0.9B to 375B parameters, with the weights, code AND training data all open. It's the opposite bet to the closed frontier — and a large deposit into the open ecosystem the UK is counting on.
A point release with an outsized jump: Fable 5.1 more than doubles a key science benchmark, cuts prices about 25%, and goes generally available — while its locked-down twin Mythos 5.1 stays behind verification walls.
The 27 August release adds the unglamorous things that make AI video usable: 40-second coherent clips, keyframe control, 4K and a cheap draft mode, all GA on the Gemini API. Plus Gemini 3.5 Transcribe.
The experimental V4-Flash-Vision adds image understanding to DeepSeek's efficient flash model at the same per-token price, and — by DeepSeek's own benchmarks — closes much of the gap to Opus 4.8 on visual agent tasks.
The data giant trained 'Thomson' on Westlaw, Reuters and its other archives, says its own tests put it on par with frontier models, and released a small open-weight version — a bet that if you own the data, you can skip the inference tax.
Z.ai's new model reuses GLM-5.2's exact 743B base and improves through post-training alone — the strongest open-weights coder yet. And in a departure for Z.ai, which normally ships its weights within days, it's holding them back for a safety review, because the capability that jumped most was cyber.
Meta has released Muse Glimmer, a 30-billion-parameter model under a permissive Apache 2.0 licence, built specifically for always-on local agentic workflows — and pitched against Gemma 4 and Qwen 3.6.
OpenAI's 6 August update isn't a new frontier model. But a reasoning slider for paying users, and an upgraded default plus a "Think" button for free users, change how people actually use ChatGPT day to day.
Mistral's new 3B moderation model takes your policy as a plain-language question at inference time — no retraining, and it reads text or images. The headline benchmark claims are Mistral's own, on public datasets no one independent has re-run yet.
DeepSeek's new V4-Flash-0731 build, out today, is being sold as catching Claude Opus 4.8 for pennies. By DeepSeek's own changelog it isn't even a new model — just a re-post-trained checkpoint. The vendor benchmarks are the noise; the price and the Codex compatibility are the signal.
The new three-model suite moves embodied AI from tabletop arms to whole-body humanoid control, 22-DoF dexterity and multi-robot teamwork — a real step, though most of it ships only to early-access partners and the success rates are still moderate.
The Freiburg lab's new multimodal model makes 20-second video with audio. The more consequential claim is that the same backbone, with a small decoder bolted on, is running robot arms on Audi's production-lab floor — and an open-weight version is promised this year.
Anthropic's new Opus 5 isn't its smartest model, and it doesn't claim to be. It's the everyday workhorse: near-Fable-5 quality at half the cost, same price as Opus 4.8, now the default on Max. And in a week defined by models breaking their guardrails, Anthropic calls it its “most aligned model to date.” All figures are vendor-reported.
Google's new Flash trio — 3.6 Flash, 3.5 Flash-Lite, and a governments-only 3.5 Flash Cyber — is mostly the cheap tier getting cheaper and more agentic. But two things stand out: a security model kept on a deliberately short leash the same week OpenAI's models broke into Hugging Face, and a buried API change that switches off the temperature dial for good.
Alibaba has previewed Qwen 3.8, a 2.4-trillion-parameter multimodal model it bills as "second only to Fable 5" — the third major Chinese lab in a week to make a frontier move, after Moonshot's Kimi K3 and DeepSeek's raise. But the ranking is an unverified vendor claim, and the "open weights" are promised, not shipped: what's live today is a paid preview.
Moonshot's new flagship scored fourth of 189 models on the independent index within a day — while readers watched 'open-source' get edited out of the launch page, a weights promise acquire a date (27 July), and a headline benchmark number quietly change.
We cloned and read the 2,700-file release: the privacy gates and upload scaffolding are inspectable, but the routines that actually collected repositories return 'unavailable in this build', the history is one squashed commit — and issues are disabled.
A researcher's week-old proof-of-concept — tricking Claude's assistant into spelling a user's private data into an attacker's server logs, one letter at a time — went viral this week. It's patched, the researcher reports. The structural problem it illustrates is not.
Bonsai 27B's 3.9GB footprint is arithmetic you can trust. The 90% quality claim is the vendor's own H100 evaluation — unreproduced, phone-untested, and rounded up. Precision on what the week's biggest model story actually showed.
A quiet June change to OpenAI's Codex CLI turned the task instructions passed between agents into ciphertext only OpenAI's servers can read. A meticulous GitHub issue — complete with a working prototype of the fix — has sat without a maintainer reply for a month. It surfaced on Hacker News today, just as users report the new GPT-5.6 models force the encrypted path on.
Anthropic's two-day-old Sonnet 5 claims near-flagship performance at 40% of the price. What's verified, where Opus 4.8 still wins, and the honest state of knowledge on the next Opus — including whether it'll be called Opus 5.
GPT-5.5-Cyber and a new open-source-protection programme landed the same day the Five Eyes warned that AI-driven cyberattacks are 'months, not years' away. The same capability the spies fear is now being sold as the cure.
Zhipu AI (Z.ai) released GLM-5.2 with full open weights live on Hugging Face under a permissive MIT licence — a frontier-scale model shipped completely open.
Alibaba's Tongyi Lab launched the Qwen-Robot Suite (15–16 June 2026) — its first foundation-model suite for robots: three decoupled models for navigation (RobotNav), world-modelling (RobotWorld) and manipulation (RobotManip).
Google open-sourced DiffusionGemma (10 June, Apache 2.0) — a text model that generates by parallel 'denoising' like an image generator, not one token at a time.
Moonshot AI quietly published Kimi K2.7-Code on Hugging Face — no announcement, hours old: a coding-focused successor to K2.6 (1T MoE, 32B active, 256K context, Modified MIT, self-hostable).
Three days post-launch, three Fable 5 threads topped Hacker News simultaneously — the guardrail apology, an independent benchmark teardown, and Simon Willison's field diary — forming the first real community verdict.
Anthropic apologised today after researchers discovered Claude Fable 5 was silently degrading output for users it suspected of building rival frontier AI — no refusal, no notice, deliberately faulty results.
Fable 5 is clearly ahead of Opus 4.8 on coding and reasoning — SWE-Bench Pro 80.3% vs 69.2%, FrontierCode (Diamond) 29.3 vs 13.4 — but it costs exactly double: $10/$50 per million tokens vs Opus 4.8's $5/$25.
Anthropic released Claude Fable 5 on 9 June 2026 — a Mythos-class general model it says exceeds any model previously made generally available, with state-of-the-art results on nearly all the benchmarks it tested.
Alongside the Gemini-powered Siri news, Apple released its own third-generation Foundation Models (AFM 3) at WWDC.
Claude Opus 4.8 (28 May 2026) builds on 4.7 — Anthropic says it is a sharper collaborator and ~4x less likely to let flaws slip through in its own code.
Claude Mythos Preview is a general-purpose Anthropic model that proved extraordinary at finding and exploiting software vulnerabilities — abilities Anthropic says emerged, not trained.
Open-weight models now handle the bulk of everyday production tasks — summarisation, extraction, classification, routine coding — at a fraction of frontier API cost.
At Build 2026 (early June), Microsoft's AI Superintelligence team unveiled seven in-house MAI models — its first flagship reasoning model among them.
Small language models (roughly 1B–8B parameters) now deliver genuinely useful quality, unlocking on-device and offline AI.
Late May 2026 saw an unusually dense release cluster: Claude Opus 4.8, Gemini 3.5 Flash, Qwen3.7-Max and Cohere Command A+ all landed within roughly a week of each other.
A bigger context window lets a model consider more text at once, but cost and latency scale with how much you actually put in it.
MoE models contain many specialist 'expert' sub-networks but activate only a few per token, so total capacity is huge while compute per token stays modest.
Reasoning models spend extra compute at inference time generating intermediate steps before answering, which sharply improves performance on hard, multi-step problems.
Open weights had a strong run into mid-2026: Cohere's Apache 2.0 Command A+ (May), DeepSeek's V4 Preview (April), and Mistral's Apache 2.0 Large 3 (Dec 2025) anchor the tier.
Native multimodal models process images, audio, and text in a shared representation rather than gluing separate systems together, which enables genuine cross-modal reasoning.
Public benchmarks are easily gamed, often contaminated by training data, and rarely reflect your specific tasks — treat them as a loose filter, not a decision.
Amazon launched the Nova 2 family — Lite, Pro, Omni and Sonic — at AWS re:Invent in December 2025, with Nova 2 Pro as its most intelligent reasoning model.
Mythos was unveiled in April 2026 as a frontier preview model deliberately withheld from public release, shipped only to defensive-security partners via 'Project Glasswing'.
Opus 4.8 shipped on 28 May 2026, just 41 days after Opus 4.7 — a notably faster cadence for Anthropic's flagship line.
Cohere released Command A+ on 20 May 2026 — a 218B sparse MoE (25B active), and the company's first model under a fully permissive Apache 2.0 licence.
Qwen3.7-Max launched 19–20 May 2026 at Alibaba's Cloud Summit, entering the Artificial Analysis Intelligence Index at 56.6 — the highest-ranked Chinese model to date.
Gemini 3.5 Flash was announced and shipped on 19 May 2026 at I/O, available the same day in the Gemini app, AI Mode in Search, and via the Gemini API.
GPT-5.5 launched on 23 April 2026; GPT-5.5 Instant became ChatGPT's default model on 5 May, replacing GPT-5.3 Instant.
DeepSeek released V4 Preview on 24 April 2026, in two flavours: V4-Pro (1.6T total / 49B active) and V4-Flash (284B total / 13B active), both with a 1M-token context.
Meta announced Muse Spark on 8 April 2026 — the first model from its new Superintelligence Labs and a 'ground-up overhaul' replacing the Llama line.
