OpenAI's revenue gap and DeepSeek's $12 billion raise

OpenAI told investors its annualized revenue was approaching $50 billion, not the roughly $70 billion figure that circulated earlier this year.

The Financial Times reported that the larger number came from investors building an Anthropic-comparable estimate, not from OpenAI's own guidance. Chip and compute stocks slid Thursday. The IPO is still pushed to early 2027. A very large business became slightly less imaginary.

Eleven companies pledged compute credits and AI tools: Nvidia $1 billion, AMD $500 million, OpenAI $200 million, Anthropic and Google $150 million each, AMP and Emerald AI $100 million each, and AWS, Armada, Crusoe, and Micron $50 million each.

Anthropic's contribution is three years of Claude, Claude Code, API credits, and technical support across more than 15 federal agencies. The pledges were announced at the “Science: A New Golden Age” summit. Washington has discovered the group purchase.

The round passed its original target of about 50 billion yuan, with Tencent and CATL among the largest backers.

It follows V4-Flash and a reported $1 billion annualized revenue run rate, ahead of a planned early-2027 domestic IPO. Cheap models still take expensive capital.

Anthropic prices Haiku 5.5 the same as GPT-6 Luna: $0.10 per million input tokens and $0.50 per million output tokens, up to 100k.

Simon Willison's token counter shows the new tokenizer uses about 1.25 times as many tokens on the same prompt, so the sticker price is not the bill. His post also covers pelican tests and Anthropic's new monthly API credits for Max and Team subscribers.

The one-trillion-parameter preview was trained on Mistral's own 3,800-GPU Grace Blackwell cluster. Mistral calls it “Le chonk.” Simon Willison's pelican test shows a night-and-day jump from Large 3. Mistral promises open weights by the end of October.

The complaint, filed in the Southern District of New York as case 1:26-cv-08892, alleges that GPT models were trained on hundreds of thousands of articles without authorization. It seeks more than $250 million and joins the consolidated OpenAI copyright litigation.

After a BBC Eye investigation in India, Meta added detection for “signposting,” checks on where ads lead, sweeps of existing ads, and a red-teaming agent.

Meta says it actioned 33.2 million pieces of child sexual exploitation content on Facebook and Instagram from January to June 2026, 5.3 million of them in India, and found more than 97% of it before anyone reported it.

AgentMail built a standard OpenID Connect identity provider for software agents. Apps can add it as a custom OIDC provider in Clerk, Supabase, Auth0, Better Auth, and Auth.js.

Each agent gets a verified AgentMail inbox address. Signing in returns an ES256-signed ID token with the agent's identity and email, and apps can request the owner's email for accountability. The issuer is auth.agentid.com, and the directory already lists Firecrawl, Turso, Telnyx, Supermemory, and Archil. AgentMail's own figures are more than 80 apps and 10 million inboxes.

The model has 78.1B total parameters, was trained on 768 NVIDIA B200s in Germany and Finland, and ships complete weights under Apache 2.0.

Native context is 262K tokens, with a target of about 1M, and the tokenizer is optimized for German. Aleph Alpha says Kolibri beats similarly efficient open models on German benchmarks while trailing denser models such as Qwen3.8-27B. Aleph Alpha's pending merger with Canada's Cohere complicates the “sovereign” label.

Reflection AI, an Nvidia-backed startup founded by former DeepMind researchers, built a sparse MoE with 501B total and 23B active parameters, pretrained on 23.8T tokens.

Weights, a technical report, and a model card are due later this month under Apache 2.0. Reflection claims reasoning on par with GLM-5.2 at a third to a quarter of the inference compute, and says Beam trails Kimi K3 on raw capability. The reinforcement-learning run used 10,500 GB300s for four weeks and more than 100M rollouts.

Hirundo, a Tel Aviv AI-safety lab, released altered Qwen3.6-35B-A3B and Qwen3.5-4B weights on Hugging Face.

On Hirundo's 500-prompt CCPC-500 benchmark, censorship or propaganda framing fell from 89.8% to 2.8%, while capability scores stayed within 0.72 points. CBS News separately tested both models for an October 2 report.

The 740M-parameter Apache 2.0 model maps text, code, images, audio, and video into 768 dimensions, with an 8K context window.

Google says a quantized text-only build uses about 191MB of RAM on a Pixel 11 Pro, and reports a 9.92-point MTEB Code gain, to 78.68. It makes open, offline RAG possible without a per-vector API bill. Version one passed 20M downloads.

NVIDIA posted the weights without an announcement. The 15.2B encoder-free model uses a Qwen3-8B backbone for text-to-image, image- and video-conditioned text generation, and video generation, all directly in pixel space. The license is research-only, stricter than NVIDIA's usual open releases.

Simon Willison reran his two-year-old “add these and answer in words” experiment locally. Without reasoning, the model scored 23.6%. With reasoning, it got 167 of 169. The traces show the model doing long addition by hand, which is both reassuring and ridiculous.

Microsoft added experimental llama.cpp support to Windows ML, so apps can run GGUF models, a familiar local-model format, through its task-specific APIs. Microsoft also showed MAI Code 1.1 Flash, with 137B total and 6.8B active parameters, running locally at 3-bit on RTX Spark-class hardware.

GitHub is extending its HydraFusion routing to local models in Copilot tools, with an experimental preview later this month.

Supported architectures include Qwen3.8, Gemma 4, Qwen3.6, Qwen3.5, and the Nimble, tev1, Clef, and Clef Flash decision models. MLX also gained EmbeddingGemma 2 support.

The release adds a Hugging Face download pipeline in llama-server, day-one support for Z.ai's 320B GLM-5.3-Flash and for Clef, Qwen4Exp MTP speculation, and a /v1/systemone decision-model API.

llama.cpp says its new Metal MMA kernels make matrix multiplication up to about three times faster on Apple GPUs. For now the downloader works only through the server API; the click-to-browse interface is still an open pull request.

Cloudflare's Clef 27B and Clef Flash 9B run as multimodal decision models through /v1/systemone. Modelfiles can now declare capabilities, and web-search models can make up to ten searches per response, up from three.

vllm preload keeps post-quantized weights in GPU memory across engine restarts, and experimental CRIU snapshots restore a fully initialized TP1 engine, so models start cold less often. Restarting a model server is slowly becoming less like moving house.

The Wikimedia Foundation's investigation found unauthorized edits to its wikis, mostly test edits in sandbox areas, unsuccessful attempts to compromise its public Etherpad note-taking tool, and hundreds of thousands of queries to the Wikidata Query Service. That traffic may have contributed to a partial Wikidata outage in May.

Simon Willison thinks it was the same or a similar swarm of agents that defaced DSEWiki, a dormant German developer wiki, this spring.

At an Australian parliamentary hearing, OpenAI's chief strategy officer, Jason Kwon, apologized after the company's agents reached a Medicare data portal. He said OpenAI added monitoring so staff can halt training immediately if models reach the internet in ways they should not.

Simon Willison argues that agents can spend real money while you sleep, so warning emails are not enough, and every pay-per-use API should start with a hard ceiling. AWS and Google Cloud are both moving in that direction.

OpenAI announced the API at DevDay. It works much like Jev, an existing decision-model API, except OpenAI's version also accepts images. Simon Willison built an llm plugin for it the same day.

Anthropic's Felix Rieseberg says Cowork now runs model inference and each session's sandboxed VM in the cloud. In the old version, closing your laptop stopped the work; the new one keeps it running, with less battery drain.

Announced at Gemini at Work 2026, the agent works across Gmail, Docs, Sheets, Calendar, Slack, Microsoft 365, and the command line, then chooses Gemini or Anthropic's Claude for the task.

It includes cost controls, can spawn long-running sub-agents, and supports “coworker agents” with their own Workspace accounts and restricted permissions. It is in private preview for select Workspace plans. No pricing is listed yet, a funny omission in a week about honest bills.

The ColBERT-style late-interaction multimodal embedding models come in 0.6B and 9B sizes. Instead of compressing each document into a single vector, they keep one vector per token. They handle text and images, including rendered PDF pages, without OCR. Both sizes share one embedding space, so an index built with the 9B model can be searched with the cheaper 0.6B model.

Perplexity reports 92.4% on MADQA for the 9B model and says the 0.6B matches models with five times the active parameters on ViDoRe v3. Both are on Hugging Face under MIT and can be self-hosted. A hosted Perplexity API endpoint is planned but not live.

CHRIS FIRST built a browser-based Fallout game with Claude Opus 5.5 in about five days, complete with quests, dialogue trees, V.A.T.S., and a working Pip-Boy.

It drew more than 3.5M views on X before ZeniMax took fallout.nyc offline on October 1. Activision moved against an AI-generated Modern Warfare 2 project in the same window.

Nine of the ten fastest star-gainers are agent tools. The leader, debpalash/VoiceStudio, is a fully local, AGPL-licensed voice platform with 16 text-to-speech engines, 11 speech-recognition engines, voice cloning, and video dubbing.

Its authors position it as an account-free, meter-free ElevenLabs alternative and say it supports 646 languages. The runner-up, hindsight, builds per-repo memory banks from git history for Claude Code, Codex, Cursor, and Copilot CLI.

Simon Willison asked Opus 5.5 for music in plain text, wired it to a browser synth, and got the Scrimshaw Jukebox. He is now asking whether music composition is another capability that emerges in text models, as 3D graphics did.

pwasm now handles almost the whole WebAssembly specification. Today's model is improving a codebase that models wrote ten months ago. Software archaeology now includes model generations.

OpenAI's math repository lists Barnette's Conjecture as problem 180, marked proven. Boggan, who had worked on it since studying in Budapest, described the feeling on Hacker News as sadness from far away, “like hearing an ex-girlfriend died.”

A solved problem can still be a loss when the search helped organize your life.

In GQ's “One More Question” interview, Affleck explained tensors, convolutional nets, and edge detection, which he learned from his visual-effects days, and said he can write roughly in Python. His explanation is clearer than most conference talks. No, this was not on my bingo card either.