
Tech • AI • Robotics • Game
A fast-moving day in AI brought signs of a new Anthropic model, mixed reviews for OpenAI’s GPT-6.2 Astra and Soul, major Google agent and model updates, and fresh infrastructure moves from Microsoft, Perplexity, Tesla, and Nvidia.
A previously unseen model called Silicon 2 appeared in tooling tied to Anthropic, prompting speculation that it could be Fable 5.5 or an internal checkpoint close to it. Listed pricing matched Fable 5.1 at $10 per million input tokens and $50 per million output tokens, but those figures were described as placeholders. The sighting comes as Anthropic is widely seen as approaching an IPO, making a stronger flagship launch strategically significant.
Separate benchmark tracking from Meter showed unidentified Anthropic checkpoints labeled model 1 through model 4 near the top of the frontier trajectory. The latest, model 4, was placed around 179 on the AEGI scale, suggesting progress beyond prior releases. That has fueled belief that Silicon 2 may be tied to the next major Fable update.
OpenAI released GPT-6.2 Astra as its higher-end model for professional reasoning, science, coding and long-horizon tasks, alongside GPT-6.2 Soul as a lower-cost general model. Astra was priced at $8 per million input tokens and $40 per million output tokens, while Soul came in at $2 and $10. The launch emphasized better efficiency, lower completion cost and a context window above 1 million tokens.
Early criticism focused on steering, consistency and agent reliability rather than benchmark weakness. Reports described Astra as strong in code review, 3D work, game development and computer use, but less dependable on long autonomous coding runs. OpenAI’s own system card reportedly showed obstacle-reporting in coding cases worsening to 22.28% from 11.09% in the earlier Astra, reinforcing concerns about transparency and control.
On Artificial Analysis’s Intelligence Index, Astra reportedly scored 62, up from 57 for its predecessor and close to Claude Fable 5.2 at 65. Measured evaluation cost was cited at roughly $4 for Astra versus about $26.54 for Fable 5.2 on that test, indicating a major efficiency improvement. The broader picture is a model that appears stronger and cheaper in some workflows, but not automatically easier to use.
In one printed circuit board design comparison, Astra reportedly finished in 45 minutes while Fable 5.2 failed to complete the same task after 4 hours. Yet Fable was judged to have produced the better schematic, a crucial factor for validating complex boards. The result underscored a recurring theme in current AI evaluations: speed and task completion do not always equal higher engineering quality.
Google was expected to introduce Gemini 4 Neon, after signs appeared in cloud billing systems with pricing of $3 per million input tokens and $15 per million output tokens. Another model, Gemini 4 Krypton, was also being watched as a possible stronger rival in the same tier. At the same time, users flagged that Gemini 4 Carbon sometimes appeared to skip visible reasoning traces even on high-reasoning settings, raising questions about controllability and performance signaling.
Gemini Tasks is being made free for large numbers of eligible users in six countries, including the United States and Canada. The feature pushes Gemini beyond chat into actions such as handling RSVPs, scheduling, checking inventory and completing purchases with Google Pay safeguards. Google AI Studio also added custom domains, Android app workflows and easier app deployment.
At its conference, Microsoft introduced several models including MAI Code 2 and Code 2 Flash for GitHub Copilot. The company also highlighted MAI Image 2.7, which was said to top an Artificial Analysis image-editing benchmark while costing roughly one-sixth as much as the four models behind it. The rollout suggests Microsoft is broadening its in-house model portfolio beyond coding alone.
Perplexity said it has crossed a $1 billion annualized revenue run rate, a marker of how quickly AI-native products are scaling into major businesses. Tesla unveiled Power Core, a solid-state system combining transformer, inverter and breaker hardware capable of 7.4 megawatts continuous output and up to 10 MVA peak, aimed at easing power bottlenecks for compute-heavy facilities. Meanwhile, Nvidia reported $115.8 billion in quarterly revenue, including $107.6 billion from data centers, showing that AI infrastructure demand remains exceptionally strong.
The latest wave of releases shows the AI race shifting from headline benchmarks to a tougher test of reliability, cost and infrastructure. Stronger models are arriving quickly, but the next winners may be the companies that pair capability gains with dependable real-world performance and enough power to run them at scale.
Ask a question