Three threads run through today’s issue.
The first is open models with price tags attached. Ricursive says its Ripple-65B cost $32 million, Cognition’s SWE-2 Flash costs a fifth of its flagship, and Tencent’s Hy4 can now operate a computer.
The second is results you can check. OpenAI paid $1.2 million for an AI to find 948 Linux root bugs with patches, Axiom turned a 7,500-page math text into 16 million lines of machine-checked Lean code, and DeepMind posted the files behind its AlphaEvolve math records.
The third is the cost of running AI. Moonshot’s Kimi K3.1 Pro holds 2.8 trillion parameters but uses 104 billion at a time, and Together AI says agents can retune its servers daily for up to 2.6 times the speed.
Open Models, New Price Tags: Who Is Selling What
Three labs shipped capable models this week, each with a cost or a catch attached.
- A chip-design startup says it built a top open AI model for $32 million. Ricursive Intelligence released Ripple-65B Preview under Apache 2.0 and scored 80.52 on TerminalBench 2.1, ahead of several much larger open models. It trails rivals on SWE-bench Verified and MCP Atlas, and the $32 million cost is the company’s own figure.
- Cognition’s new cheap coding AI is fast, but you can’t download it yet. SWE-2 Flash costs a fifth of SWE-2, runs three times faster, and is free in Devin Desktop and the Devin command-line tool through 30 November. Cognition is holding back the weights and the benchmark comparisons are its own.
- Tencent’s Hy4 can now read images and operate a computer. The full launch adds image understanding and computer use, and Tencent says the hallucination rate fell from 12.95% to 7.2% against the August preview. The weights and safety documents are only promised for the coming days, and the scores come from Tencent’s own tests.
Show the Receipts: Bugs, Proofs and Fake-Photo Detectors
Four releases where the real question is whether anyone else can verify the result.
- OpenAI spent $1.2 million to get AI to find 948 ways to hack Linux. OpenAI says an unreleased cyber model turned 2,308 suspected kernel flaws into 948 unique root-access bugs, each with a patch, for an API bill of $1,213,430. The $66.3 million bounty value is OpenAI’s own pricing of its work.
- AI agents turned a 7,500-page math reference into checked code. Axiom says its agents rebuilt the Stacks Project as about 16 million lines of Lean, a language whose checker rejects any step that does not follow, and caught statements in the original that did not hold. The claim of biggest ever is Axiom’s alone.
- DeepMind publishes the code behind its math-solving AI, with a catch. The AlphaEvolve repository includes a minimal working version and the files behind its claimed math records, so skeptics can rerun the evidence. The best matrix multiplication bound still needs a large file that readers must download and upload themselves.
- Pangram’s new detector says it can spot fake photos and fake videos. Pangram’s image model marks where in a photo the AI-made part sits and reached 99.96% accuracy on 1,200 fakes from 12 generators. Pangram built that test itself and does not name the rival detectors it compared against.
Cheaper to Run: Smaller Bills Under the Hood
Three ideas for making each answer cost less, from the model’s design to the servers behind it.
- Kimi’s big open AI model is now downloadable, and here is what’s inside. Moonshot published the weights for Kimi K3.1 Pro, which has 2.8 trillion parameters but activates only 104 billion per token. Memory needs follow the total, not the active count, and the benchmark scores still await outside testing.
- Cartesia’s AI skips the word-chopping step and learns from raw text. Cartesia says swapping the tokenizer for a model that learns its own text boundaries beat a standard design at every size tested, with a 1.28 times compute gain on its validation data. The results come from Cartesia’s own research post.
- Together AI lets software agents retune its model servers daily. Together Compiler keeps one engine running and swaps each candidate tweak in, which Together measures at 9 seconds instead of a 467 second cold start. The company says retuning takes under 24 hours and yields speeds up to 2.6 times faster.
Today’s Quick Hits
- OpenAI’s image update adds GIFs and draws 50% faster. Images 2.6 brings animated GIFs to ChatGPT and Codex on every plan, plus two developer models, Flare and Sunburst. The 50% speed gain and six-second render time are OpenAI’s own figures.
- Conductor, the multi-agent coding app, is joining Meta and closing its cloud. Conductor Cloud goes read-only on 13 November and closes on 20 November, with a Transfer to local option for workspaces. The Mac app gets support for new models and serious bug fixes through early 2027.
- The product chief behind Claude Code has moved to OpenAI. Cat Wu, who ran Claude Code at Anthropic, joins OpenAI’s core products team and will report to Tibo Sottiaux. OpenAI confirmed the hire, and Anthropic declined to comment on her departure.
- Paper says AI models can work fine with a much shorter memory. A report from Peter Belcak and eleven co-authors says compacting a session often costs frontier models little accuracy while avoiding a cost gap of up to 7x. Only the abstract was available, so the evidence is unseen.
- Perplexity’s Decisions API can now check the web before it answers. Adding web_grounding mode lets the probability-returning API search for evidence first. In Perplexity’s own 328-claim test, accuracy rose from 56.4% without grounding to 97.3% with it.