Issue 11: Anthropic files to go public
Anthropic filed its S-1 on 26-10-2026 and plans a Nasdaq listing under the ticker ANT. The filing reports 16.3 billion dollars of revenue in the first half of 2026 at 85 percent gross margin, nearly half of it through the Bedrock, Vertex and Azure marketplaces. Gemini 4 Carbon took first place on Artificial Analysis’s Intelligence Index, and trusted cyber defenders get it before wider developer access. Moonshot’s Kimi K3.1 Pro is live, with open weights promised for 04-11-2026. An audit found Claude Code sending background agents’ full transcripts about every 30 seconds to write a status line, at 16.4 percent of one team’s Claude Code spend in a week.
- Gemini 4 Carbon tops evals, opens to defendersArtificial Analysis ranks Google DeepMind's model first on its Intelligence Index at 59, though its hallucination rate rose from Argon's 4 percent to 11 percent. Trusted cyber defenders get it first through the Fairwind Program; wider developer access is promised as soon as possible.
- Anthropic files its S-1Revenue rose from 41 million dollars in 2023 to 16.3 billion in the first half of 2026 at 85 percent gross margin; Q2 operating profit was 974 million dollars. Cloud marketplaces carried 48 percent of first-half revenue; their fees sit outside that margin.
- Kimi K3.1 Pro lands, open weights promisedMoonshot's 2.8-trillion-parameter, 104B-active model costs 2 and 10 dollars per million input and output tokens, with open weights promised for 04-11-2026; a Highspeed version runs 2 to 3 times faster, per Moonshot, at twice that. It ranks fourth on Artificial Analysis's Intelligence Index, at 55.
- Spawn: a game world built by Fable agentsPlayers reshape Spawn's world by prompting its agent, Savi. By Jacob Sansbury's account, a swarm of Fable-powered agents builds it for over 3 million dollars in tokens a month, 30 times payroll, shipping to production hourly; Simon Willison doubts the style suits higher-stakes software.
- Elorian tests if models see past a decoyElorian and Mercor's DecoyCount tests whether models count objects or lean on visual shortcuts; in the paper, the best frontier model, GPT-6 Astra, scored 43.17 percent. Elorian's CEO says Glimpse, which generates and edits images while reasoning, scored 50.48 percent in one internal run.
- Custom voices with Qwen3-TTSBuilding a dungeon crawler with his kids, Armin Ronacher picked Qwen3-TTS for its narrator: speaker embeddings are 2,048 numbers you can blend, and 28 seconds of his conference talk gave a halfway convincing clone that he notes could trick someone's grandmother over the phone.
- Claude Code burns tokens on a status lineAbout every 30 seconds, Claude Code sends background agents' full transcripts to their model for a 3 to 5 word status line; no setting stops it. One team measured 16.4 percent of a week's spend, others 18 to 26 percent; a Bun preload blocks it.
- MCP Tasks and Elicitations come to ClaudeClaude is getting support for MCP Tasks, which let it hand plugins long-running jobs such as deep research and hear back when they finish, and Elicitations, which let plugins ask users for input mid-task, such as confirming a refund.
- GPT-6.1 Terra reaches Codex and the APIOpenAI says the model matches GPT-6 Astra on Agents' Last Exam at a tenth the per-token price and finishes tasks over 3 times faster on average, with standard API prices 50 percent below GPT-5.6 Terra's; Astra or GPT-6.1 Sol remain recommended for complex coding.
- API credits for ChatGPT subscribersEligible ChatGPT Pro and Business subscribers can claim up to 500 dollars a month in free API credits, activated from ChatGPT or the API Platform by linking an API organization they own.