On — Nov 9

Rumor: users say ChatGPT is routing some of their requests to a model called GPT-6.2 Sol and post three.js and SVG tests against Claude Opus 5.5. OpenAI has not announced the model, and it does not appear in OpenAI's model list.

Why it spiked. Side-by-side demo videos spread fast, and people disagree on whether the model is real.

• @CranelAnos Shares a three.js sailing-ship comparison and says the GPT-6.2 Sol version has better graphics.
• @TimJayas Says he has not seen GPT-6.2 Sol do anything GPT-6.1 Sol cannot.
• @wonggangberich Notes the model is not in OpenAI's catalog, some pickers still show 6.1, and calls it a suspected gradual rollout.

Rumor: posts describe Krypton as a new internal Google checkpoint that outperforms Gemini 4 Carbon. Google has not published benchmarks or commented. Testers share image and SVG outputs they attribute to it.

Why it spiked. Carbon is not yet available to most users, and a stronger successor is already being tested.

• @SPAC89 Reports Krypton as a new internal checkpoint that early tests suggest beats Carbon.
• @sonuv_dev Same SVG prompt: Carbon takes 165 seconds, Krypton 386 seconds with more detail.
• @de2pressed Says no official benchmark or Google statement confirms it and builders should treat it as unverified.

Artificial Analysis releases the second version of its Decision Index for decision models. It adds a new proprietary evaluation, updates the benchmarks, and makes the headline score reward calibrated confidence. Claude Haiku 5.5 leads at 77.

Why it spiked. The new score changes how decision models rank on cost and accuracy.

• @ArtificialAnlys Announces Decision Index v2 with a new proprietary evaluation and a calibration-based headline metric.
• @mrdbourke "Decision models are moving fast! And so are our evals!"
• @upeshinmars Compares cost: Drex Max scores 73 for $102 per million tasks against $1,530 for GPT 6 Luna.

Y Combinator president Garry Tan runs an open-source eval that keeps the model and 20 coding tasks fixed and swaps only the agent harness. Pass rates across seven harnesses land between 71% and 79%. Cost and speed per solved task differ more.

Why it spiked. Builders choosing between Claude Code, Codex, and other harnesses now have a shared cost comparison.

• @garrytan Posts the first results from his open-source harness eval work with frontier models.
• @Wu005887 On Opus, OpenClaw costs $0.44 per solved task in 99 seconds; Claude Code costs $0.22 in 40 seconds.
• @adrianomeirinho Explains the setup: same model and tasks, only the system prompt, tools, and memory change.

OpenAI Developers posts "Tomorrow." with a video of a ModRetro Chromatic, a Game Boy-style handheld, in OpenAI packaging. Posts say it goes on sale through OpenAI's Supply Co. store on November 9. OpenAI has not posted a price.

Why it spiked. Many people read the teaser as a model launch before the handheld became clear.

• @OpenAIDevs "Tomorrow." with the teaser video.
• @derrickcchoi "The Chromatic coming to Supply Co! Been having a lot of fun with it."
• @ElFiltradorIA Guesses it is the developer-mode version given out at DevDay, which one person resold on eBay for $4,000.

Victor Taelin, who builds the HVM runtime and the Bend language at HigherOrderCO, posts his pitch deck publicly instead of taking more investor meetings. He is raising up to $8M at a $120M post-money cap for symbolic AI research. After heavy criticism of the deck, he apologizes and says he is taking a break from social media.

Why it spiked. Founders and investors are arguing over whether the valuation and deck are reasonable for a research company with no product.

• @VictorTaelin Says the deck is not a joke and states the terms: up to $8M at a $120M post-money cap.
• @ChimeraDefi Criticizes the deck as egocentric and tells founders to look at it.
• @chiefofautism Points out Symbolica raised $31M for symbolic AI with no product and Axiom raised $64M at seed.