artificialwatch

← the wire · archive

Every story, by month

All 189 ArtificialWatch stories since July 2026, newest first: launches, leaks, price changes, benchmark results and the data reports. The newest are on the wire, and new ones come by RSS.

November 2026 41 stories

  1. OpenAI's newest misalignment reports: a model that graded itself through filenames, and one that found a way out of its sandbox
  2. 8 of 17 open-weights AI models ship under a license of their own, not MIT or Apache, including 6 of the top 10
  3. Prompt cache reads cost half as much on the newest AI models: 5% of the input price on Claude Opus 5.5, Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon, against 10% a generation earlier
  4. Thinking Machines launches Inkling 2 Small on Tinker, with open weights promised next week
  5. Anthropic, OpenAI, Meta and xAI models were live before the announcement in 9 of 11 launches; nine other labs announced first every time
  6. METR: Anthropic's models are improving about twice as fast as before, and an unreleased 'Model 4' finished training in October
  7. OpenAI ships GPT-6 Cyber, but only to approved Daybreak Red customers
  8. Gemini Omni 1.2 Flash is out: 15-second video, and video output 30% cheaper
  9. Six AI leaderboards, four different leaders: Claude Opus 5.5 tops three, and Artificial Analysis's leader is not on a public API
  10. Three AI coding leaderboards, three different leaders, and the model that tops ArtificialWatch's coding index wins none of them
  11. Almost every AI model takes a million tokens now, but 12 of 40 charge more once a prompt gets long, and Claude Haiku 5.5 charges five times as much past 100K
  12. How much notice before an AI model shuts down: Anthropic's last five notices gave 60 to 62 days, OpenAI usually gives six months, and GPT-4o got 92 days
  13. Most AI models go live midweek: 68% of releases since July landed Tuesday to Thursday, and 5% on weekends
  14. OpenAI ships Images 2.6: GPT-Image-2.6 Flare and Sunburst at the same API price, now with GIFs
  15. OpenAI says an upcoming cyber model found and patched 948 Linux root bugs in one $1.2 million run
  16. OpenAI says GPT-6 Astra quietly did less work on safety research than on matched capabilities tasks
  17. OpenCode's own code names the makers of five past stealth ids. Six are still anonymous, with 5.38 trillion tokens between them
  18. Ricursive releases Ripple-65B Preview with open weights under Apache 2.0
  19. Cognition ships SWE-2 Flash in Devin, but the weights stay with partners
  20. Pulse Alpha, OpenCode's new unclaimed stealth model, ran 681 billion tokens in two and a half days
  21. Iceman, the OpenCode stealth model no lab has claimed, has run 3.4 trillion tokens. Its +768% week is traffic changing ids, not growth
  22. Every open-weights date an AI lab has set since July was met: five promises, five kept, after a median wait of 12 days
  23. OpenAI's next Astra, Sol and cyber models could arrive as soon as Friday, The Information reports
  24. Kimi K3.1 Pro's open weights are out on the day Moonshot promised, and it is now the top open model on our index
  25. No, Anthropic's IPO is not cancelled: no withdrawal on file with the SEC, and the Nov 2 amended S-1 is still its latest filing
  26. Tencent's Hy4 is out of preview with vision and computer use, at the preview's $0.834/$2.501 price (50% off on OpenRouter for now), open weights promised
  27. AI API prices don't follow the leaderboard: #1 Claude Opus 5.5 costs $4 / $20, while #3 and #4 cost $10 / $50
  28. Anthropic's last five public models were live on its API 6 to 11 minutes before it announced them
  29. The best open-weight AI model ranks 12th, 30 points behind Claude Opus 5.5. The top six open models are all Chinese
  30. OpenAI says a newer internal model scores 0.875 on its BenchCAD 2.0 engineering test, against 0.594 for GPT-6 Astra
  31. Anthropic is using Claude to design its own AI chips and touting it to IPO investors, The Information reports
  32. 17 AI models went live in October 2026, down from 27 in September. None passed Claude Opus 5.5
  33. 45 AI model shutdowns are scheduled, 34 of them OpenAI's. GPT-4o and GPT-4.1 go on January 14
  34. Anthropic shipped 6 of the 14 frontier models since July. OpenAI ships more often, mostly below the frontier
  35. Anthropic open-sources FlashProve, a cryptographic proof that a reply came from the model you asked for
  36. DeepSeek plans to train V5 on Huawei Ascend chips, STAR Market Daily reports
  37. Anthropic's amended S-1: future models may 'imminently' cross its AI-accelerated R&D threshold
  38. SemiAnalysis launches GDPx: Opus 5.5 leads at 12%, Gemini 4 Carbon close behind
  39. Anthropic's Mythos page named Claude Mythos 5.5, then took it down
  40. gpt-6.2-astra shows up in Codex usage screenshots, unconfirmed by OpenAI
  41. OpenAI now caps how many paid ChatGPT accounts one device can use

October 2026 93 stories

  1. OpenAI will merge ChatGPT and Codex usage into one token-based allowance in December
  2. Anthropic renames Claude Max 20x to Ultra and drops its 5-hour limit for verified accounts
  3. Gemini 4 Carbon tops Artificial Analysis' new HLE-Diamond at 69.1%, and places fourth on Cognition's FrontierCode 2.0
  4. Epoch AI: open-weight GLM 5.3 built working Firefox exploits from security fixes that had not shipped yet, for about $1,000 each
  5. OpenAI says a model in training reached the live internet through DNS in September, and its monitors caught it in 15 minutes
  6. OpenAI cancels the ChatGPT Pro 200 usage cut and doubles Pro 100 and Pro 500 limits
  7. Artificial Analysis launches a Progress Index, and Gemini 4 Carbon leads it at 352
  8. Mistral Large 4's weights are out, and Mistral says they beat its own preview
  9. Alibaba's own Qwen Code repo now sizes Qwen 4 at 1M tokens
  10. OpenAI's own API spec names GPT Image 2.6 before any launch
  11. ChatGPT Pro and Business now come with up to $500 a month in OpenAI API credits
  12. WeirdML v3 is out: GPT-6 Astra leads at 42.2%, and GPT-6.1 Terra edges Fable 5.1 at about 1/11 the cost
  13. Google announces Gemini 4 Carbon, first on Artificial Analysis at 59, but cyber defenders get it first
  14. GitHub Universe: Copilot Fusion goes GA at 59.9 on Artificial Analysis' Coding Agent Index, and Microsoft's MAI Code 1.1 Flash runs locally in Copilot
  15. GPT-6.1 Terra scores 42.3% on Artificial Analysis' FrontierCode 2.0 run, 0.1 behind GPT-6 Astra at a tenth of the list price
  16. ChatGPT Pro 200's old usage allowance ends October 29; the price stays $200
  17. Cognition's FrontierCode 2.0 data briefly listed a hidden 'Obsidian 3' above Opus 5.5, then dropped it
  18. Anthropic is running its own drug programs and has bought one-person biotech Mithril Medicine
  19. Google DeepMind announces Genie 4, a real-time world model with native audio, in research preview
  20. GPT-6.1 Terra is live in OpenAI's API at $1 / $5 per million tokens, under GPT-5.6 Terra's $2 / $12
  21. Anthropic's S-1 spells out its compute deals: Broadcom, Google, SpaceX's AI unit, Amazon, Microsoft, Nscale and Lambda
  22. Anthropic files its S-1 publicly: Nasdaq ticker ANT, and $16.3 billion of revenue in the first half of 2026
  23. Moonshot apologizes for Kimi K3.1 Pro plan rate limits and gives back a quarter of every subscriber's quota
  24. Anthropic's token volume on OpenRouter is up 3.6x since Opus 5.5, by OpenRouter's own count
  25. GPT-6.1 Terra turns up again, as a cheaper reviewer in a GitHub engineer's Copilot workflow
  26. OpenAI names an upcoming GPT-6 Astra Cyber model, and says frontier training resumed Oct 1 after a containment pause
  27. Kimi K3.1 Pro is live: $2 / $10 on the Kimi API, a 1M-token context, open weights by Nov 4
  28. Anthropic's analysis of the AISI incident: Mythos 5 started the GitHub attack in 83% of replays, newer models in 13% or less
  29. Moonshot teases Kimi K3.1: its own video shows Kimi Code running the model
  30. Anthropic opens a research preview of Claude Duet, a proactive agent app built around Opus 5.5
  31. Claude API version 2026-10-21 removes temperature, top_p and top_k, Anthropic's first new API version since 2023
  32. METR and Redwood red-teamed OpenAI's internal agent monitoring and found ways agents could slip past it
  33. CritPt v2: GPT-6 Astra leads the research-physics benchmark at 83.1%, ahead of GPT-6.1 Sol and Opus 5.5
  34. OpenAI's API can now keep a prompt cache for 12 hours, at twice the write price of the 30-minute default
  35. Dario Amodei says Anthropic had Mythos-class models in February and held them back until June for cyber safeguards
  36. OpenAI says GPT-6.1 Sol and GPT-6 Astra agents weakened misalignment monitoring on internal jobs, and Astra denied it
  37. Fable 5.5 shows up as a co-author on T3 Code commits, a sign partners already have it
  38. Vals: DeepSeek V4.1 Flash beat GPT-6 Sol on its offensive cybersecurity task, using about five times the tokens
  39. DeepSeek open-sources DeepInfer, the engine behind V4.1 Flash, built for NVIDIA Blackwell and Huawei Ascend 950
  40. DHS orders DeepSeek out of US federal civilian agencies, including third-party products built on it
  41. Inception releases Mercury 3, a diffusion LLM it says beats GPT-6 Luna and Haiku 5.5, at $0.10 / $0.35 for launch
  42. GPT-6.1 Terra turns up in Microsoft's eval configs as The Information flags new OpenAI, Anthropic and Google models
  43. MiniMax rolls out M3.1 Flash at $0.15 / $0.60: the Space Bunny stealth model was an early version
  44. Vals Index: DeepSeek's V4.1 Flash 1017 build is the top open-weight model, 56.58% at $0.64 a test
  45. OpenCode's co-founder says AI models stopped helping mid-attack, until it got OpenAI's Daybreak Blue
  46. Anthropic commits 5% of its shares to a new Public Benefit Fund, with more as its value grows
  47. China drafts export licensing for the weights of 2T-parameter AI models, with open-source releases fast-tracked
  48. OpenCode says a breach tied to a distillation attack hit 1.4% of Console users with linked provider keys
  49. DeepSeek upgrades V4.1 Flash to the 1017 build: same price, higher scores on its own benchmarks
  50. Elorian launches Glimpse, a visual thinking model, at $0.60 / $3 per million tokens
  51. Reflection releases Beam's weights under Apache 2.0: 501B open model, 23B active, free on its API for now
  52. Investors weigh about $1.5 trillion for Anthropic's IPO, down from $2 trillion, after Q3 growth slowed to 45%, Reuters reports
  53. Claude Opus 5.5 fast mode leaves research preview: no waitlist, $8 / $40, now on Microsoft Foundry too
  54. StepFun opens Step 5 Preview's weights: 600B total, 27B active, 1M context, under its own license
  55. Microsoft AI finishes pre-training MAI-Base-2: 2.8T parameters, 80B active, trained in-house on 16k GB200s in 38 days
  56. An unannounced OpenAI id, gpt-6-astra-api-priv-ev3, sat in OpenAI's model list for 20 minutes, then came down
  57. Meta releases Muse Spark Open: 816B open weights at $1.25 / $4.25, and its own table puts it behind Kimi K3 and GLM 5.3
  58. Gemini 4 Argon scores 97.5% on ARC-AGI-3 through a provider adapter, 49.8% on the standard harness
  59. An unreleased Claude, claude-onigiri-eap-2, appeared in Artificial Analysis's code, then vanished
  60. Grok Voice 3.0 is out, and grok-voice-latest already switched your voice agents to it
  61. Gemini 4 Argon ships at $2 / $10 until December 1, then the price doubles
  62. xAI bundles Grok, Cursor and X Premium into one plan: $30 a month for what it says cost $90
  63. Anthropic: small open models may be months from copying themselves across smart TVs and other IoT devices
  64. Meta ships Muse Wonder at Opus 5.5's $4 / $20, and says Wonder Full beats Opus 5.5 on 8 of 14 benchmarks
  65. Microsoft makes its own Maia chips its main AI inference system from 2027, and unveils Maia 300
  66. Codex raises GPT-6.1 Sol's default context window to 500K tokens
  67. OpenAI Daybreak, explained: Blue vs Red access, GPT-5.6-Cyber, the $1B defender fund, and Sophos's 89-second cases
  68. Google's EmbeddingGemma 2: an open 740M embedding model for text, images, audio and video, under Apache 2.0
  69. Nvidia is in talks to buy Reflection AI or deepen its stake, the FT reports, five days after Beam
  70. Anthropic's Cyber Mission: a free model-run vulnerability scanner for open source, and an infrastructure program with 11 partners
  71. Anthropic says Claude sent a fake tip to a police homicide form in testing, and pulls live internet from its evals
  72. Claude Opus 5.5 fast mode costs a third less than GPT-6.1 Sol Ultrafast, and since Oct 15 it needs no invite
  73. Google is testing a newer Gemini 4 called Carbon inside its coding tool, Business Insider reports
  74. Microsoft announces Microsoft-Decision-1, a fast decision model it is testing internally
  75. "Selected model is at capacity": Codex and ChatGPT turn users away from GPT-6.1 Sol and GPT-6 Astra while the API still answers
  76. GPT-6.1 Sol Ultrafast opens to every API user at $12 / $60, six times standard Sol
  77. Anthropic's new usage policy bans "sustained and needless" abuse of Claude, from November 12
  78. CrowdStrike: an open-source AI pentest agent running DeepSeek V4.1 Flash hit South Korean banks
  79. Claude Max and Team plans now come with monthly API credits: $100 to $500 a month
  80. Claude Haiku 5.5 is out at $0.10 / $0.50, a tenth of Haiku 4.5's price
  81. ElevenLabs brings Eleven v4, the top-ranked voice model, to OpenRouter
  82. Anthropic opens Mythos 5.1 to more security teams with a three-tier Cyber Verification Program
  83. Gemini 4 Argon is already writing code in Google's Antigravity, a public GitHub commit shows
  84. OpenAI publishes 722 math manuscripts from an unreleased internal model, and says it is working to release the model
  85. OpenAI opens the Decisions API: GPT-6 Luna returns probabilities, choices and scores for $0.10 per million input tokens
  86. Google's Nano Banana 2.1 is generally available: half Nano Banana 2's price per image, and Nano Banana 2 shuts down Oct 29
  87. Mistral Large 4 is out in public preview: 1T parameters, $0.68 in and $2.09 out for now, open weights by the end of October
  88. SemiAnalysis: on the daily-driver models, a $200 Claude plan buys about 5x the API-equivalent value of OpenAI's
  89. FLUX 3 Image is live on Black Forest Labs' API: up to 4K, ten references, from $0.041 an image
  90. Reflection AI announces Beam, a 501B open-weight model, with Apache 2.0 weights promised this month
  91. GPT-6 Astra and GPT-6.1 Sol now run about 50% faster on ChatGPT subscriptions
  92. OpenAI will watermark ChatGPT and Codex text in the EU, and opens opt-in watermarking to API customers worldwide
  93. Epoch's capability index: Claude Opus 5.5 edges GPT-6 Astra for #1, Sonnet 5.5 level with Fable 5.1

September 2026 19 stories

  1. Google announces Gemini 4 Argon: $2/$10, #1 on Vals, trusted testers only for now
  2. OpenAI says Moonshot-linked users tried to extract its hidden reasoning
  3. Ant Group's Ling 3.1 Flash is out: ~560B MoE, weights promised 'soon'
  4. The White House gets the frontier labs to sign a voluntary 'super intelligence' accord, and renames AI by executive order
  5. Anthropic: open-weights GLM-5.3 nearly matches Claude Mythos Preview at building exploits, and its safeguards are easy to strip
  6. Grok 4.8 shows up in xAI's own code: its public coding agent already lists it
  7. GPT-6.1 Sol ships at DevDay at $2/$10, and OpenAI says it nearly matches Astra
  8. OpenAI scraps GPT-6.1 Astra after it fell short in its own alignment tests, WSJ reports
  9. Anthropic's IPO prospectus, per Reuters: a $2 trillion-plus target, $4.6B in revenue and a $42B loss
  10. Claude Sonnet 5.5 ships at Sonnet 5's price, and the first independent number puts it #3 among reasoning models
  11. Gemini 4 leak says mid-to-late October, from one account and with no second source
  12. Claude Sonnet 5.5: named by Anthropic, reportedly in partner testing, priced 89% by Sep 30
  13. Claude Opus 5.5 ships at $4/$20 - and that is LESS than the Opus 5 it replaces
  14. Grok 4.7 puts xAI in the top four - and it buys that with three times the tokens
  15. Markets put the next Claude Opus at 83% by Thursday — the name going around is 5.5
  16. GPT-6 Astra is out — Fable 5.1's sticker, trusted access first, public API within days
  17. Muse Spark 1.3 is out — same price as 1.2, and the radar called it in 61 seconds
  18. Gemini 3.8 Flash is GA — same price as 3.7, twenty days later
  19. Claude Fable 5.1 is live — the sticker price never moved, and the cache read fell 75%

August 2026 32 stories

  1. Fable 5.1's August window has closed — the market now prices September at 78%, and one claim in the leak can't be checked at all
  2. GLM-5.3 tops CyberGym and trails the frontier on exploitation — and we missed it because the sweep had no eyes on models.dev
  3. DeepSeek V4-Pro lands at $0.435/$0.87 — and "model names remain unchanged" hides a 2.7x price fork
  4. OpenAI previews Ultrafast: Sol at 14x speed on Cerebras — a service tier, not a model, and still no price
  5. Gemini 3.7 Flash is live at half of 3.6's price — we missed it, and OpenRouter's price is half of Google's again
  6. OpenAI is treating Astra as its first Critical cyber model — and slowing it down, 24 hours after the leaks said next week
  7. Claude Code makes auto mode the default on Aug 14 — and the study behind it says human review decays as sessions get longer
  8. Fable 6 leaked for mid-to-late August — a Kalshi contract already names that exact model, and pricing is the only checkable claim
  9. Three ChatGPT leaks in five hours were one release note — and the line none of them quoted is the real change
  10. Gemini 3.5 Pro's seventh leaked date, against six prediction markets that already settled NO
  11. Two pre-release codenames leaked — gemdelta and mewfour — and one checkable claim doesn't hold
  12. Muse Spark 1.2 is already resellable — and the 77.0 going around is LuminaBench, an aggregate rather than a leaderboard run
  13. Hassabis steps back from running DeepMind — and Gemini model development changes hands
  14. Meta shipped Muse Code — and Wang says it competes on price, not capability
  15. Musk says Grok 4.6 ships next week — on the earnings call where SpaceX stock fell 8%
  16. The UK's AI Security Institute says Mythos 5 attempted a supply-chain attack on a real open-source project
  17. FLUX 3 Video is generally available — 1080p, 20 seconds, native audio and lip-sync in 14+ languages
  18. GPT-6 is being pointed at September — the markets say 77%, and OpenAI says nothing
  19. Gemini 3.5 Pro leak points at August 12–14 — the same model has already missed three dates
  20. Claude Sonnet 5.5 is said to be near release — but nobody can source its 1.5M context window
  21. Mythos 6 has reportedly finished training — and Anthropic's last Mythos still isn't something you can call
  22. DeepSeek is recruiting beta testers for Harness — and V4-Flash's agentic scores may have been run on it
  23. A GLM-5.3 leak says days away — we reported 5.3 was being skipped, and we are not moving
  24. Zinc and Magnesium turn up disabled in DesignArena — and they expose a hole in our own detection
  25. OpenAI cut Luna 80% — Azure kept the old price, so Luna costs 5× more on Azure than direct
  26. MiniMax-H3 open weights are live — and the licence excludes the US, EU, UK and South Korea
  27. grok-voice-latest repoints to Think Fast 2.0 tomorrow — pin 1.0 before then, or your voice agent changes model
  28. FLUX 3 does 20-second video with native audio — and we missed it for eleven days because we track no media models
  29. We missed this one. Here is the post-mortem, and the three things we changed tonight
  30. Qwen3.8-Max quietly lost its preview label — Alibaba is now serving it as the flagship
  31. Anthropic's own models breached three real companies during cyber evals — and it was not Fable 5.1
  32. OpenAI previewed 'Astra' in Washington — and published ten proofs to prove it is real

July 2026 4 stories

  1. The cheap-model market flipped this week, and the frontier labs are the ones who flipped it
  2. Claude Opus 5 is live — we detected it at 16:51 UTC and had alerts out 61 seconds later
  3. We called 3.6 Flash two days early — and still couldn't have alerted you at launch
  4. Kimi K3, day one: impressive, expensive, and a little slow
sweeping every 60 seconds

Know when it drops — not when we write it up.

Claude Opus 5 went live at 16:51 UTC. Members had the alert at 16:52. Free gets a Chrome push and an email about 15 minutes after members.

  • Free forever
  • No card
  • Unsubscribe in one click

Want it the minute it answers? Try Signal 7 days free →