klarapelhe.rs/en/ai-news
OpenAI Finally Drops GPT‑6.2 Astra and Sol but Claude Fable 5.2 Holds Its Lead
OpenAI finally launched GPT‑6.2 Astra and Sol on November 17, eight days after Claude Fable 5.2 debuted. Both improved on the Artificial Analysis Intelligence Index, where Fable still leads. On the Coding Agent Index, Astra held steady and Sol dropped three points.
Contents
- Astra and Sol get better general scores. Coding tells a different story
- Astra takes first place in interactive 3D
- “I do not like 6.2 Astra,” Theo wrote after early testing
- What GPT‑6.2 Sol gets done for the money
- Fable 5.2 scores higher, but Astra costs 85% less per benchmark task
- Google opens Carbon to developers and adds a cheaper Gemini
- Nvidia rules out Jensen Huang’s return after alleged Chinese intelligence link
- Useful links
Key takeaways
- OpenAI launched GPT‑6.2 Astra and Sol on November 17. Sol matched the previous Astra’s overall score at about 15% of the cost per task. It scored below the previous Sol on Artificial Analysis’s coding benchmark.
- Astra averaged $4 per task at maximum reasoning effort. Claude Fable 5.2 scored higher on the same benchmark and averaged $26.54 per task.
- Google opened Gemini 4 Carbon to paying developers and released Hydrogen for faster everyday tasks.
- Jensen Huang resigned as Nvidia CEO on November 12 after reports about his relationship with a former executive suspected of ties to Chinese intelligence. Trump backed Huang’s return, and Nvidia’s chair said on November 17 that the board would keep looking for a successor.
Astra and Sol get better general scores. Coding tells a different story
OpenAI released GPT‑6.2 Astra for complex work and GPT‑6.2 Sol for everyday tasks. Both can inspect a codebase, edit files and run tests. OpenAI said early testers got cleaner code from Astra with fewer revisions. The company urged users to judge both models by what they could get done.19
On Artificial Analysis’s Intelligence Index, which tests skills like reasoning and knowledge, Astra gained five points. Only Claude Fable 5.2 scored higher. Sol gained two points to reach 57, matching the previous Astra at a much lower cost per task.2
Artificial Analysis also tested both models in Codex at maximum reasoning effort. Astra matched its previous coding score, and Sol fell three points compared with GPT‑6.1 Sol. On a test that asked questions about an existing codebase, Sol dropped 14 points because it left out important details more often.210
Astra takes first place in interactive 3D
Give two models the same brief for an interactive 3D scene, hide their names and let a judge play with both. That is how Artificial Analysis’s AA‑3D test works. Astra earned the highest preference rating, 1,318 Elo, ahead of Fable 5.2 at 1,271 and Sol at 1,259. Elo measures relative preference across pairs. A model can win people over with a playable world even when another model scores higher across reasoning and knowledge tasks.12
Computer use asks something different of the model. It has to operate software to get from a brief to a usable result. In one OpenAI demonstration, Astra took photos of Chicago’s Marina City towers as the starting point for a paper lamp, then revised the design in computer-aided design software. A person assembled the lamp. In another, Astra modelled a car, wrote a Python simulation of how its tyres and suspension interact, and animated the result in Blender. The animation showed what the proposed design would do under the model’s assumptions. Checking that it predicts the behaviour of an actual car would require physical measurements.1
Sol shares some of that visual range, as its third-place 3D result suggests. OpenAI describes it as an everyday model that can combine text, images and tools such as web search in one task. For Astra, the company gives the more demanding coding example of tracing a bug across services and testing a fix. The independent coding results above show why the precise job matters: explaining an existing implementation and producing a new interactive scene draw on different skills.112
“I do not like 6.2 Astra,” Theo wrote after early testing
Theo, who builds T3 Code, said he would choose Claude Opus 5.5 over Astra for his own development work. OpenAI had fixed some of the problems he reported, but his experience still kept him from recommending Astra. He said it sometimes stopped in the middle of a long task to ask about details of its own implementation. After long runs, he used Claude Opus 5.5 to make sense of Astra’s account of what it had changed. He praised its game creation and computer use, and credited it with finding a way to route incoming traffic that he expected would save him money.12
Dan Shipper reported that his team at Every tested the models for several weeks and preferred Fable 5.2 overall. He liked Astra’s first design choices, but found it difficult to steer successive designs away from a recognisable house style. On coding builds the team often had to restart work Astra had stopped early. In writing, Shipper found that it sometimes compressed an answer until the reasoning was difficult to follow. Theo had the opposite complaint about length: Astra’s explanations buried the point in implementation detail. Both accounts describe extra work for the person receiving the answer.1312
Ethan Mollick supplied a small example of design repetition. Given freedom to choose the setting for a spy puzzle game, his Astra chose “Bellwether Station.” Theo had already described seeing the same unusual name in games produced from different prompts. The recurrence is a detail two testers could compare directly. How frequently Astra repeats creative choices across users remains unknown.1412
OpenAI’s leaders judged the launch by other criteria. President and co-founder Greg Brockman foregrounded professional work and scientific research, arguing that the models could deliver both at a lower cost per task. Artificial Analysis lends some support to the science claim: Astra led its terminal-based scientific-workflow test and ranked behind only Fable on the hard subset of its physics evaluation. Those results help explain how the same release could look strong to a researcher and frustrating to someone waiting for a finished software build.152
Sam Altman acknowledged on launch day that OpenAI still had work to do to make its models easier to work with. He said that would be a priority for upcoming releases.16
What GPT‑6.2 Sol gets done for the money
In the standard API, Astra costs $8 per million input tokens and $40 per million output tokens, both down 20% from GPT‑6 Astra. Tokens are the units of text the service processes and bills for. Sol stays at $2 and $10. Its lower cost on the general index therefore matters even to teams whose per-token price for Sol has stayed the same.12
Thibault Sottiaux, who works on Codex at OpenAI, said Sol Ultrafast was already his everyday choice. The launch gives Pro 200 subscribers access to that faster mode for Sol. Pro 500 includes it for both models. Standard Astra and Sol are also available in ChatGPT Work and Codex on Plus, Pro, Enterprise and Edu, with administrators enabling access in eligible organisational workspaces.117
Altman also singled out Sol Ultrafast and the amount of work the models could do with fewer tokens. Sottiaux gave that efficiency a practical meaning: shorter waits, lower task costs and more work before reaching a plan limit. These were launch-day assessments from OpenAI. The early accounts from Every and Theo add a cost that token billing misses: a person may still have to restart an agent or decipher its handover.18191312
Fable 5.2 scores higher, but Astra costs 85% less per benchmark task
Anthropic released Claude Fable 5.2 and Mythos 5.2 on November 9. They share an underlying model. Fable has safeguards that allow broader distribution through Anthropic and cloud platforms. Organisations seeking Mythos for cybersecurity or biomedical research can apply through verification programmes. For a team choosing a model for ordinary development work, Fable is the relevant comparison.4
At their standard API rates, both models charge $8 per million uncached input tokens and $40 per million output tokens. The bill for a task can be very different. At maximum reasoning effort, Artificial Analysis measured a weighted average of $4.00 per Intelligence Index task for Astra (62 points) and $26.54 for Fable (65 points). Astra cost about 85% less on this workload. The gap depends on the task and reasoning setting.3 Theo’s account offers a different comparison again. The Claude model he used to make sense of Astra’s work was Opus 5.5.12412
One example behind Anthropic’s confidence in Fable is a migration of an internal production service from Python to Rust. Anthropic says agents using an early version of Fable 5.2 carried out the migration and the service subsequently started about 26 times faster. That figure measures startup time for one service.4
Fable’s visual work also drew praise from an early tester. Unity creative technologist Sam Zhang described building 3D scenes from scratch and reconstructing experiences from visual references. What impressed Zhang was Fable’s ability to keep refining difficult game-development tasks across iterations. Anthropic selected the testimonial for its announcement. It helps explain Fable’s appeal to a developer refining an existing scene, even as judges preferred Astra’s finished experiences in AA‑3D.42
Google opens Carbon to developers and adds a cheaper Gemini
Google’s October preview of Gemini 4 Carbon was restricted to selected partners. On November 10 it opened the model to paying developers through the Gemini API. Through November 30, a million input tokens cost $2 and a million output tokens $10. Afterward, the rates rise to $4 and $20. A developer budgeting a deployment on the introductory price would therefore underestimate December’s token bill by half at the same volume and mix.5
Gemini 4 Hydrogen, released November 12, costs $1.50 per million input tokens and $7.50 per million output tokens across its million-token context window, the amount of material it can handle in one request. Developers can choose among four reasoning levels. Google reports that at the default high setting it finished its test tasks 2.5 times as fast as Gemini 3.8 Flash. Free Gemini app users can also encounter Hydrogen when the app automatically selects it for a prompt.6
Google illustrated Carbon’s visual and coding capabilities with a browser game set on its Mountain View campus. Carbon coordinated agents that built 3D assets in Blender, assembled the game in Three.js and added sound. For less playful work, Google and Artificial Analysis devised a test of whether agents could reason over charts, schematics and other visual files. Google reported Carbon in second place. Google helped design that evaluation.5
Google also tested using Carbon and Hydrogen together in its Antigravity development environment. Its Orchestrator switches models within a session, drawing on Carbon for harder decisions and Hydrogen for quicker execution. On Terminal-Bench 4.0, Google reported nearly Carbon’s task success while finishing 19% faster and consuming 24% less quota on average. That result measures the combined system and its routing decisions.6
Nvidia rules out Jensen Huang’s return after alleged Chinese intelligence link
The allegation that brought Jensen Huang’s leadership of Nvidia into question concerned Alice Huang, a former executive at the Singapore computing firm Megaspeed with whom he had an affair. Reuters reported that some US authorities suspected she had ties to Chinese intelligence. Beijing denied the allegation. On November 8, Nvidia’s board ordered an independent investigation into reports that its founder had been targeted by a foreign intelligence service. Whether Alice Huang acted for Beijing remained unresolved as of November 18.82021
On November 12, Jensen Huang resigned as president, CEO and a director. Chief financial officer Colette Kress took over as interim CEO. Within days, his possible return had divided the Trump administration. Treasury Secretary Scott Bessent called for sweeping changes at Nvidia on November 15. The next day, Trump said he would welcome Huang’s reinstatement. Elon Musk also backed his return.78
On November 17, Huang publicly offered to come back if directors believed that would serve the company, its employees, shareholders and the country. The board gave investors its answer that day. Chairman Harvey Jones said directors had unanimously accepted the resignation and had no intention of reinstating him. Kress would lead Nvidia while they searched for a permanent chief executive.822
A federal inquiry explains why the case reached beyond Nvidia’s boardroom. The US Commerce Department says it began examining possible chip diversion involving Megaspeed in 2024. Investigators requested Nvidia sales records in May 2025. The following month Nvidia put Megaspeed on its no-sell list, and US and Singaporean officials inspected the company and interviewed Alice Huang. That timeline establishes what investigators did to trace the chips. The intelligence allegation concerns a separate question about whom Alice Huang may have worked for. On November 17, Nvidia’s board created a committee to oversee national security and export compliance.202324
Useful links
- OpenAI announces Astra and Sol
- Artificial Analysis compares both generations in Codex
- Theo on Astra’s strengths and limitations
- Sam Altman addresses the experience of working with GPT‑6.2
- Tibo explains his preference for Sol Ultrafast
- Dan Shipper on coding and design at Every
- Ethan Mollick on a recurring name in game designs
- Greg Brockman on the cost of professional work and research
Sources
- OpenAI. Introducing GPT-6.2 Astra and Sol. 17 November 2026.
- Artificial Analysis. OpenAI has the #2 model in the Intelligence Index with the release of GPT-6.2 Astra and GPT-6.2 Sol. 17 November 2026.
- Artificial Analysis. Release comparison tool.
- Anthropic. Introducing Claude Fable 5.2 and Claude Mythos 5.2. 9 November 2026.
- Google. Gemini 4 Carbon is now available. 10 November 2026.
- Google. Introducing Gemini 4 Hydrogen. 12 November 2026.
- Nvidia. NVIDIA Announces CEO Transition. 12 November 2026.
- Associated Press. Nvidia chairman says board does not intend to reinstate Jensen Huang as CEO. 17 November 2026.
- OpenAI. GPT-6.2 launch announcement. 17 November 2026.
- Artificial Analysis. GPT-6.2 coding-agent results. 17 November 2026.
- OpenAI. Model release notes for November 17. 17 November 2026.
- Theo. Why I hate GPT-6.2 Astra. 17 November 2026.
- Dan Shipper. Every’s early tests of writing, coding and design. 17 November 2026.
- Ethan Mollick. A familiar setting in an Astra-generated game. 17 November 2026.
- Greg Brockman. GPT‑6.2 Astra and Sol for professional work and research. 17 November 2026.
- Sam Altman. Working with the new models. 17 November 2026.
- Thibault Sottiaux. Sol Ultrafast on Pro 200. 17 November 2026.
- Sam Altman. Token efficiency and Sol Ultrafast. 17 November 2026.
- Thibault Sottiaux. Task speed, cost and usage limits. 17 November 2026.
- Reuters. US releases details of Megaspeed probe, says investigation began in 2024. 17 November 2026.
- Nvidia. Statement from NVIDIA Board of Directors. 8 November 2026.
- Jensen Huang / Associated Press. Jensen Huang’s statement on a possible return. 17 November 2026.
- U.S. Department of Commerce / Department of Justice. Joint statement on the Megaspeed investigation. 16 November 2026.
- Nvidia. NVIDIA Board of Directors Establishes National Security and Export Control Oversight Committee. 17 November 2026.