OpenAI shipped GPT-6 Astra on September 3, one day after Google released Gemini 3.8 Flash. Both models’ largest improvements sit in coding and autonomous agent performance, a shared optimization target that two competing labs reached independently and nearly simultaneously. The same week, Kilo Code shipped a rebuilt JetBrains plugin that can run either model, or any of 500 others, inside the same workflow. Together, the three releases mark a widening separation between the model layer and the orchestration layer that runs it.

Optimized autonomy

OpenAI released GPT-6 Astra last Wednesday at $10 per million input tokens and $50 per million output tokens, pricing it at 2.5 times the rate of GPT-5.6 Sol. The model carries a 1.05-million-token context window, supports up to 128,000 output tokens, and ships with native computer use, a hosted shell, and MCP support, a toolkit built for models that operate browsers, fill forms, and chain tools across multi-step workflows. OpenAI’s launch demos showed Astra formatting legal contracts, building a 3D game, and booking appointments as autonomous sequences, navigating between applications and chaining actions without returning control between steps.

The benchmark profile makes the optimization target visible. Astra scores 97.6% on FrontierMath Tier 4, up from Sol’s 80.5%, and 57.9% on Terminal-Bench 4.0, up from 37.3%. On the Artificial Analysis Intelligence Index, which measures general reasoning across a broader task set, the gap between Astra and Sol is 61.2 to 60.9. That same independent assessment found that Astra performs on par with Anthropic’s Fable 5 at lower cost on coding-agent work, and on par with Sol on general intelligence. OpenAI president Greg Brockman told reporters that Astra marks the beginning of “the AGI era,” though the general-intelligence numbers suggest a more incremental advance than the agent benchmarks.

A push to the frontier

Google shipped Gemini 3.8 Flash the day before Astra, on September 2, marking its third Flash-class release in six weeks. The three iterations, 3.6 Flash on July 21, 3.7 Flash on August 13, and 3.8 Flash on September 2, all hold the same introductory pricing of $0.75 per million input tokens and $3.75 per million output tokens through the end of 2026, while each ratchets up coding and agentic benchmark scores. Google rolled 3.8 Flash into the Gemini app, AI Mode in Search, and Gemini in Sheets for Pro and Ultra subscribers on the same day as the API launch, making the agentic improvements immediately available across consumer and enterprise surfaces.

Each lab paired its flagship release with a security-specialized variant. Astra carries OpenAI’s first “Critical” cybersecurity designation under its Preparedness Framework, a classification indicating that the model can discover and exploit unknown vulnerabilities without step-by-step human guidance. OpenAI reported that the model found two previously unknown vulnerabilities during its own evaluation, a result that prompted the company to gate advanced cyber capabilities behind its Daybreak program for vetted security researchers. Gemini 3.8 Flash Cyber approaches the same domain from the defensive side, restricted to government and critical-infrastructure operators through Google’s Fairwind Program, where Google reports that it produces 2.6 times more correct patches than larger commercial models. Cybersecurity, like coding-agent performance, has become a shared optimization target across frontier labs.

The JetBrains buffet

Kilo Code, the open-source coding agent acquired by Anaconda in July, shipped a rebuilt native JetBrains plugin last week. Rewritten from the ground up in Kotlin, the plugin supports parallel agents in isolated Git worktrees and connects to more than 500 models at provider rates with no markup. Astra, Gemini 3.8 Flash, Fable 5.1, or any other model with an API endpoint slots into the same development environment without reconfiguration. A developer can route different tasks to different models within a single session, choosing the frontier tier for complex architecture decisions and a cheaper model for mechanical implementation.

When frontier models converge on the same capabilities at similar price points, the tools that orchestrate them absorb differentiation that the models once carried on their own. A developer choosing between Astra at $10/$50 and Gemini 3.8 Flash at $0.75/$3.75 for a coding task faces a trade-off between capability ceiling and cost per token, with both models designed for the same category of work. Choosing which IDE plugin, terminal agent, or cloud harness to build around carries more practical consequence for daily work than the model selection, because the harness determines environment, integrations, and the feedback loops that shape how code gets written.

Critical metrics

Models released a week apart already target the same benchmarks, and each successive release will narrow the remaining gaps. The orchestration layer, meanwhile, accumulates project-specific context, workflow integrations, and team configurations that make switching costs asymmetric. Replacing a model inside a harness is a configuration change that takes effect on the next API call. Replacing the harness itself means rebuilding those integrations from scratch. For model vendors competing on benchmarks that narrow with each release, the strongest lock-in in the developer toolchain may belong to the layer they do not control.

Subscribe now

AI Central

Convergent Evolution