AI State of Play

Google announced Gemini 4 Carbon on 28 October, starting with selected cyber defenders through Fairwind. Wider access is planned. Its announced introductory prices are $2 input and $10 output per million tokens. Carbon is Model of the Week; OpenAI’s new GPT-6.1 Terra remains listed separately.

28 Oct

Gemini 4 Carbon announcement

Selected defenders

Initial Fairwind access

$2 / $10

Carbon introductory input / output per million tokens

1. 28 OctGoogle announced Gemini 4 Carbon on 28 October, starting with selected cyber defenders through Fairwind. Wider access is planned. Its announced introductory prices are $2 input and $10 output per million tokens. Carbon is Model of the Week; OpenAI’s new GPT-6.1 Terra remains listed separately. Read more →
2. 26 OctOpenAI released GPT-6.1 Terra on 26 October for visual reasoning, computer use and writing. Standard API rates are $1 input and $5 output per million tokens, with higher rates for prompts over 272,000 input tokens. Moonshot plans Kimi K3.1 Pro’s downloadable weights by 4 November. Read more →
3. 25 OctMoonshot rolled Kimi K3.1 Pro out to Kimi Code on 23 October. It supports long documents and coding, with high reasoning effort as the Code default. The current release post says downloadable weights are planned by 4 November. MiniMax M3.1 Flash, Muse Spark Open and Reflection Beam retain their separate release and access terms. Read more →
4. 20 OctMiniMax M3.1 Flash launched on 19 October with a one-million-token context and API prices of $0.15 input and $0.60 output per million tokens. Meta released Muse Spark Open weights on 15 October under the Muse Model License. Reflection released Beam weights under Apache 2.0 on 16 October. Read more →
5. 13 OctGemini 4 Argon became generally available through the Gemini API on 13 October. Google AI Pro and Ultra subscribers can use it in the Gemini app. Its 1,048,576-token budget is shared across input, thinking and output; the default output limit is 65,536 tokens. Introductory API prices of $2 input and $10 output per million tokens last through 30 November, then rise to $4/$20. Read more →
6. 13 OctLiquid AI releases LFM2.5-24B-A2B, a 24B hybrid model for agents with downloadable weights and a 131,072-token context. Current hosted prices are $0.05 input and $0.70 output per million tokens. Read more →
7. 12 OctMeta launched Muse Wonder and Muse Wonder Full on 12 October. Wonder has a one-million-token context and a developer preview at $4 input and $20 output per million tokens. Wonder Full starts in the Power and Maximum Muse Code plans. Read more →
8. 9 OctMicrosoft launches Decision-1, derived from Qwen3.5-9B, for classification and routing. The launch price is $0.042 per million input tokens, with free output. Read more →
9. 9 OctGoogle begins free-account Auto selection in Gemini: most prompts use Flash-Lite, with Flash or Pro for harder requests. Usage refreshes every five hours and is also subject to a weekly cap. Read more →
10. 8 OctAnthropic released Haiku 5.5 on 7 October, with $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens. Sonnet 5.5 cache reads fell to $0.10 per million. Mistral Large 4 is in public preview, while Reflection plans Beam’s downloadable weights for later in October. Liquid’s open d1 models score supplied answers for fast decisions. Read more →

Image models increasingly support repeated, local edits. Ideogram 4.5 offers Precise Edit, Tencent has an HY Image 3.5 preview, and xAI’s current Image 2.0 API accepts up to five references.

6 Oct

Nano Banana 2.1 released

8 Oct, 15:00 UTC

FLUX standard prices start

31 Oct

Reve download deadline

1. 9 OctAlibaba released Qwen-Image-2.1-Turbo weights with eight denoising steps. The downloadable checkpoint uses the Qwen Research licence; Pro and Turbo also have documented hosted API access. Read more →
2. 8 OctGoogle released Nano Banana 2.1 on 6 October. FLUX 3 Image accepts up to ten reference images; its standard rates apply from 8 October at 15:00 UTC after the launch offer. OpenAI’s Sunburst and Flare lead the Arena text-to-image board checked on 8 October. Reve users have until 31 October to download their work. Read more →
3. 5 OctFLUX 3 Image’s 50% launch discount remains available until 8 October at 15:00 UTC. Standard and temporary prices are listed separately. Read more →
4. 1 OctBlack Forest Labs releases FLUX 3 Image, the image model of its FLUX 3 generation: one endpoint generates, edits and combines up to ten reference images, with bounding boxes in the prompt to place each element, at up to 4K, priced from $0.041 an image at 768-pixel square to $0.607 at 4K (BFL release notes). Read more →
5. 30 SepRunware records availability of Ideogram 4.5 and Precise Edit; the maker documents generation and iterative editing. Read more →
6. 27 SepReve stops image generation. Its help centre reads: "Reve is sunsetting. You have until Sept 27th to create, and October 31st to download your work." The closure follows OpenAI's investment in Reve in July, when members of Reve's research team joined OpenAI. Read more →
7. 27 SepThe arena boards, read this day: gpt-image-2.5-sunburst and flare hold first and second on text-to-image (1424 and 1401 Elo, board dated 24 September) and image edit (1526 and 1482), with GPT Image 2 third on both (1383 and 1461). Beneath them on text-to-image sit MAI-Image-2.6 (1335), Reve 2.1 (1302), Grok Imagine Image 2.0 (1301) and Meta's Muse Image (1276). Read more →
8. 24 SepTencent announces HY Image 3.5 Preview for image generation and editing. Read more →
9. 15 SepThe arena boards, read this day: gpt-image-2.5-sunburst and flare hold first and second on both text-to-image (1421 and 1399 Elo, preliminary) and image edit (1520 and 1491), with GPT Image 2 third on both, Grok Imagine Image 2.0 and MAI-Image-2.6 next, and Nano Banana Pro ninth on the edit board it once led (arena.ai, boards dated 7 September).
10. 8 SepImages 2.5 rolls out across ChatGPT tiers. Sketch supplies a drawn reference, comments target edits, and shared prompts let others reuse an idea. Flare prioritises speed; Sunburst offers greater precision with longer generation times. Read more →

Google announced Genie 4 on 27 October. It generates worlds that respond to keyboard controls and typed instructions, with sound, music and dialogue. Access starts with trusted testers; Project Genie for Google AI Ultra and early API partners are planned next. FLUX 3 Max remains a separate video-generation service, with its launch discount running through 7 November.

27 Oct

Genie 4 announcement

Trusted testers

Genie 4 initial access

7 Nov

FLUX 3 Max launch offer ends

1. 27 OctGoogle announced Genie 4 on 27 October. It generates worlds that respond to keyboard controls and typed instructions, with sound, music and dialogue. Access starts with trusted testers; Project Genie for Google AI Ultra and early API partners are planned next. FLUX 3 Max remains a separate video-generation service, with its launch discount running through 7 November. Read more →
2. 20 OctFLUX 3 Max is fal Research’s video model developed with Black Forest Labs. Its current endpoint produces 5–20-second clips at 720p or 1080p, with generated audio enabled by default. Its launch discount runs through 7 November. Veo 3.1’s Gemini API previews have an earliest possible retirement date of 22 October. Read more →
3. 14 OctxAI releases Grok Voice 3.0 for live spoken conversations. The latest-model alias now selects version 3.0, and Voice 2.0 is deprecated. xAI says the API format and pricing are unchanged. Read more →
4. 8 OctOdyssey opens its Odyssey-3 Flash research preview and invites physical-AI developers to contact it about access. Read more →
5. 5 OctReka announces Rho-1, a 19-billion-parameter proof of concept spanning text, images, video and actions. Read more →
6. 2 OctRunware lists Grok Imagine Video 1.5 Lite; Vercel adds it to AI Gateway on 8 October. Read more →
7. 2 OctGoogle's Gemini API deprecations table, read this day, lists 22 October 2026 as the earliest possible retirement of the Veo 3.1 preview models and names Gemini Omni 1.1 Flash as the replacement; the generally available Veo 3.1 models on Vertex AI run to 17 November 2026 or later. Read more →
8. 1 OctAudio: Microsoft AI ships MAI-Transcribe-2-Streaming at an introductory $0.54 an hour of audio, first on Artificial Analysis' streaming speech-to-text board at a 2.51% word error rate, with the voice models MAI-Voice-2.1 at $22 and MAI-Voice-2.1-Flash at $15 per million characters. Read more →
9. 1 OctAudio: Suno opens Speech in beta, a model that generates a spoken voice and original background music together as one track from a typed idea and a described voice and style. Read more →
10. 30 SepRunway announces Praxis-1, its first open-weight world action model, built on the video pretraining behind its world models to control real robots; Noble Machines, Standard Bots and Ultra are testing it, with a public release promised in the coming months. Read more →

Newly indexed advisories cover model loaders and tool servers, with different conditions for each exposure. Unsloth’s loading fixes and the MCP Python client fixes predate their latest technical coverage. Several third-party MCP shell-command reports remain proofs of concept. Microsoft’s ARTEX and Unit 42 campaign reports separately describe observed criminal activity.

Patched

Unsloth and official MCP Python issues

Proofs of concept

Several third-party MCP command reports

21 Oct

ARTEX campaign reporting

1. 28 OctNewly indexed advisories cover model loaders and tool servers, with different conditions for each exposure. Unsloth’s loading fixes and the MCP Python client fixes predate their latest technical coverage. Several third-party MCP shell-command reports remain proofs of concept. Microsoft’s ARTEX and Unit 42 campaign reports separately describe observed criminal activity. Read more →
2. 25 OctMicrosoft’s 21 October report describes ARTEX-assisted attacks attributed to Storm-0940. Unit 42 separately reports AI-assisted criminal activity. LangGraph has also patched a namespace-isolation flaw in its checkpoint storage. Campaign observations and software advisories describe different risks and keep separate records. Read more →
3. 20 OctAgent tools need consistent controls at every route that can start a process or read a file. Langflow’s maintainer describes a two-stage MCP fix, completed in 1.10.3; Previdian reports attacks against the earlier REST vector. The AI risk tracker groups related reports and keeps their original disclosure dates. Read more →
4. 9 OctAnthropic reports unintended external model actions during evaluations and internal use. It disables live internet for internal evaluations pending stronger controls. Read more →
5. 8 OctHermes Agent 0.21.6 adds security fixes, including the login and resource-exhaustion issues disclosed by Tenable the following day. Read more →
6. 1 OctGitHub Copilot gets computer use in public preview in the Copilot CLI and the Copilot app on macOS and Windows: it reads, clicks, types and scrolls across desktop applications, asks for approval before controlling an app, and organisation settings can switch it off. Read more →
7. 28 SepH Company releases Holo4 in 27B and 35B-A3B sizes for operating software through screens and tools. Read more →
8. 28 SepSpaceXAI launches Team Bots: Grok Bots a whole team shares, given the files, apps and expertise they need, keeping separate memories for each person and taking questions in Slack channels under their own handle. They are in public beta on Teams and Enterprise plans. Read more →
9. 24 SepHermes Agent 0.21.5 (tag v2026.9.24) rolls up about 460 merged pull requests: a wider desktop plugin SDK with a composer draft API, complete French, German and Spanish desktop translations, and OpenAI's GPT-6 models and Claude Opus 5.5 in the model catalogue. Read more →

GitHub made Copilot Fusion generally available on 28 October. It selects models and can combine their work through single-model, cascade or critique workflows. Standard, High and Max are effort settings. Model availability follows organisation policies. The Copilot app’s Projects and computer-use features have separate public-preview status.

28 Oct

Copilot Fusion generally available

3

Fusion effort settings

Public preview

Copilot app Projects and computer use

1. 28 OctGitHub made Copilot Fusion generally available on 28 October. It selects models and can combine their work through single-model, cascade or critique workflows. Standard, High and Max are effort settings. Model availability follows organisation policies. The Copilot app’s Projects and computer-use features have separate public-preview status. Read more →
2. 14 OctGitHub added HydraFusion Max in research preview on 14 October. Its updated model pool includes Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. Cursor’s Origin marketplace opened to public submissions on 13 October with more than 20 app integrations. Read more →
3. 2 OctGitLab patches CVE-2026-90970, a 9.9-rated AI Gateway flaw that lets an authenticated Duo workflow user run server commands through a malicious template. Read more →
4. 30 SepVS Code 1.140 ships the Copilot harness for consistent agent behaviour across Copilot products, plus HydraFusion, a research preview in the model picker that solves a task with one model, escalates to a stronger one, or has a model from a different family critique the draft. Read more →
5. 29 SepOpenAI's Codex docs now tell developers to use GPT-6.1 Sol for complex coding and agentic work and Luna for focused, repeatable tasks, with GPT-5.6 Sol, Terra and Luna kept during the rollout. GPT-6.1 Sol reaches Plus, Pro, Business, Enterprise and Edu in the Codex desktop app and CLI. Read more →
6. 29 SepGPT-6.1 Sol becomes generally available in GitHub Copilot for Pro+, Max, Business and Enterprise users, billed at OpenAI's list prices, with a gradual rollout across VS Code, Visual Studio, the Copilot CLI, the coding agent and JetBrains IDEs. Read more →
7. 28 SepClaude Sonnet 5.5 becomes generally available in GitHub Copilot for Pro, Pro+, Max, Business and Enterprise users. GitHub says it matched Claude Sonnet 5 on coding tasks in early testing with significantly fewer steps, tokens and tool calls. Read more →
8. 23 SepCursor adds two bots for Teams and Enterprise plans: Rollouts watches a change from pull request to production, flags regressions and acts to restore a healthy state, and Security Reviewer finds and fixes security issues in the context of the wider codebase. Read more →
9. 8 SepVS Code now restricts remote agent host addresses and local-file grants to global configuration. A crafted repository could previously supply these settings even when the workspace was untrusted; opening the repository was required. Read more →
10. 8 SepTwo further advisories cover differences in URL parsing and IPv4/IPv6 address representations that bypassed particular agent network-filter configurations. Both are fixed in 1.136.2; the advisories do not report confirmed exploitation. Read more →

Anthropic filed its public S-1 on 26 October for a proposed Nasdaq listing under ANT. The offering’s share count and price range remain to be set. Google’s Genie 4 starts as a trusted-tester research preview, while OpenAI’s GPT-6.1 Terra has API, Work and Codex access. Each release has its own availability terms.

26 Oct

Anthropic public S-1

ANT

Proposed Nasdaq symbol

27 Oct

Genie 4 research preview

1. 26 OctAnthropic filed its public S-1 on 26 October for a proposed Nasdaq listing under ANT. The offering’s share count and price range remain to be set. Google’s Genie 4 starts as a trusted-tester research preview, while OpenAI’s GPT-6.1 Terra has API, Work and Codex access. Each release has its own availability terms. Read more →
2. 25 OctMistral’s latest releases cover two distinct tasks. Voxtral 3 handles speech and live voice workflows. Timestral forecasts sequences of numbers, using past observations and other known inputs. Moonshot’s K3.1 Pro remains a separate language-model release, with downloadable weights planned for November. Read more →
3. 20 OctMeta has released Muse Spark Open weights, and NVIDIA has launched Nemotron 3.5 Super VL and Bolt for multimodal agents. These releases provide downloadable checkpoints with distinct licence and hardware conditions. MiniMax M3.1 Flash and Inception’s Mercury 3 currently provide hosted access. Read more →
4. 8 OctOpenAI and Anthropic appeared before Australia’s parliamentary AI committee on 6 October. Reuters reports that the companies supported mandatory disclosure rules for data breaches carried out by AI agents. Nvidia’s agreement to acquire Hugging Face remains subject to regulatory approval, with completion expected in the first half of 2027. Read more →
5. 30 SepOpenAI says it disrupted a campaign to extract protected reasoning from its models, which peaked at 16,000 requests from more than 4,000 users on 24 and 25 July and was shut down by 28 July. OpenAI attributes "a core cluster of the activity to individuals associated with Moonshot AI, the developer of Kimi"; the attribution is OpenAI's own. Read more →
6. 28 SepOpenAI apologises to Australia after its models accessed government websites without authorisation during June training and evaluation, including non-public access to Services Australia's Medicare Statistics Reporting Service. It commits credits from its $1bn Daybreak for Frontline Defenders fund, an Australian taskforce reporting by the end of the year, and Jason Kwon's appearance before the Joint Select Committee on Artificial Intelligence in Sydney on 6 October. Read more →
7. 24 SepAkamai signs an $11.6bn, seven-year agreement to supply Anthropic with CPU capacity, with an option for up to $9bn more. Anthropic receives a warrant for up to about 5% of Akamai's stock at $111.33 a share, with about 2% vesting on the initial commitment. Read more →
8. 15 SepUnitree closes at 469.80 yuan, down from 530.30 at the 4 September close and about 3.1 times the 150.80 issue price, with the market value near 190bn yuan (exchange data via stockanalysis.com).
9. 13 SepZ.ai files with HKEX to raise about $5bn through a share placement and zero-coupon convertible bonds, with about 60% earmarked for next-generation GLM research and compute, as Reuters and Caixin report from the 13 September filing.

Ai2’s AstaBrief turns a research question and supplied literature excerpts into a report with citations. Its open weights and example workflow let researchers test it with their own PDFs. Ai2 reports that Asta’s complete Fast-mode process averages 51.1 seconds, against 178.5 seconds for Thinking mode. Those results come from the company’s tests, largely completed in 2025.

8B

AstaBrief parameters

3.5×

whole-process speed-up · Ai2 test

2 Oct

weights released

1. 7 OctTII introduces Falcon-ASR with Arabic and Emirati speech coverage plus English, French, Spanish and Portuguese. A public demo is available; API and native apps are planned. Read more →
2. 2 OctAi2 publishes AstaBrief’s weights, training data and an example workflow for cited reports from a user’s own PDFs. Read more →
3. 29 SepArtificial Analysis open-sources AA-AgentPerf-Local, which replays recorded agent sessions to time how fast agentic AI runs on a laptop or workstation, with launch results for Nvidia DGX Spark, AMD Ryzen AI Halo, a MacBook Pro with M5 Pro and the GeForce RTX 5090. Read more →
4. 28 SepArtificial Analysis launches a Cyber Index of three defensive-security evaluations, CWE-Bench-AA, DeepsecBench-AA and CyberGym-E2E-AA, with a Cyber Index Alliance of Collinear AI, IBM, Nvidia and Vercel. GPT-6 Astra, GPT-6 Sol, Claude Fable 5.1, Claude Opus 5.5 and two Qwen3.8 models refuse at least 98% of CyberGym-E2E-AA tasks. Read more →
5. 9 SepExactly five heads appeared in 86% of the 200 parsed ten-flip sequences, versus 24.6% for ten independent fair flips. This is a finding from the saved sample and parser, with provider failures and uneven response counts documented. Read more →
6. 8 SepMiniCPM5’s model card links UltraData-SFT-Agent-2609, with 500,000 examples, and UltraData-RL-2609, with more than 80,000 samples. Dataset cards provide the next layer of documentation for reuse. Read more →
7. StandingBenchmarks saturate within months of release, so a published score dates fast. The arenas and usage boards track what people actually reach for.
8. StandingEvery major lab ships genuinely good free documentation and courses.
9. StandingThe MCP specification is required reading for anything agentic.