Opus 5.5 and GPT-6 Sol and Luna land an hour apart, and the price sheet gets rewritten

GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output. That is half the price of GPT-5.6 Luna, which was already the cheapest capable model most people had used, and it landed about an hour after Anthropic shipped Claude Opus 5.5 at $4/$20, a 20% cut from Opus 5 with cache reads down 60% to $0.20. Simon Willison’s same-day writeup has the full table, and the shape of it is the story: Grok 4.7 and Xiaomi’s MiMo v2.6 came out the day before, GPT-6 Sol now sits at $2/$10 where GPT-5.6 Terra used to be, and the only models left above $4 input are Fable 5.1 and GPT-6 Astra at $10/$50. AWS Bedrock and GitHub Copilot listed Sol and Luna within the hour, both with 1M-token context, and OpenAI’s internal factuality eval says Sol makes roughly half as many mistakes as GPT-5.6 Sol. Anthropic’s launch tweet passed 17 million views by the evening, and AINews moved its own production pipeline to Opus 5.5 the same night on the strength of the writing improvements.

Two details from Willison’s testing matter more than the benchmark claims. Opus 5.5 at max effort failed his pelican-on-a-bicycle SVG prompt twice by reasoning past the 128,000-token output ceiling before emitting anything, at $2.56 and nearly 20 minutes per failure, so he now treats max as effectively unusable and runs medium as his Claude Code default. And Artificial Analysis found that at max effort Opus 5.5 costs $5.98 per Intelligence Index task against $5.86 for Opus 5, because an 80% increase in token usage eats the whole sticker discount; the “40% cheaper per task” claim holds at the medium default, not at the top. Sonnet 5.5 and Haiku 5.5 are promised within weeks, which is where the price war actually bites, since Haiku 4.5 at $1/$5 is now ten times the price of Luna.

CodeRabbit ran Opus 5.5 through its review pipeline against 80 known bug patterns from its August open-source benchmark plus 13 harder cases, and the result inverts the Opus 5 finding. Opus 5 had bought precision by catching fewer known bugs; 5.5 edged past CodeRabbit’s production baseline on coverage in the open-source set and gained on both coverage and precision in the harder set, at the cost of more comments and more tokens. The integration changes are the part to read before you upgrade: thinking is always adaptive and the API rejects requests that explicitly enable or disable it, effort is the only deliberation control, and forced tool calls are retired, so any workflow that required the next response to call a specific tool now has to check whether the call happened and handle the case where it did not. Medium effort on 5.5 matched or beat Opus 5 at high on multi-step coding using about half the tokens.

JetBrains Air folds six months of separate experiments (JetBrains Central, its CLI, cloud agents, automations, cost controls) into one product system with three named surfaces: Air in the IDEs for directing agents and verifying their output, Air Teams for delivery workflows that mix developers and autonomous agents, and Air Governance, the renamed Central, for policy, audit, and AI cost management. The framing is unusually direct for a 26-year IDE company: “the era in which the whole software development system can be contained in one window is ending,” and the system is explicitly multi-vendor, meant to hold agents JetBrains does not build. Junie stays the in-house coding agent. What to watch is whether Air Governance becomes the thing enterprises actually buy while the IDE surface competes with Copilot and Claude Code on the merits.

Google open-sourced AX, a Kubernetes-style orchestrator for autonomous agent workloads, per InfoQ. It runs on a substrate that treats each agent as a stateful actor, with a control plane exposing Kubernetes-style primitives for tasks and resources, and the headline capability is suspending and resuming agent tasks so idle phases stop burning compute. If you have been hand-rolling a worker pool that parks long-running agents between tool calls, this is the thing to evaluate before you write more of it.

Cloudflare’s Worker Previews give every Git branch a production-like environment with its own URL, variables, secrets, bindings, observability, and, importantly, its own Durable Objects and Containers, so per-branch state and migrations do not bleed across concurrent tests. A preview deploys with npx wrangler preview, inherits a base configuration you define, can be overridden per branch to point at a scratch database, and can serve on a custom domain so OAuth redirects and cookies behave as they will in production. Cloudflare pitches this as the runtime half of the Agent Development Lifecycle it proposed last week: an agent pushes a branch, tests against a real URL, reads its own logs and traces, fixes, and verifies before anything reaches production.

Meta’s Muse is the security story of the day. Ars Technica reports that Muse, “Meta’s extraordinarily privileged AI assistant,” has a serious 0-day, and a separate post from mouse.dev that reached 125 points on Hacker News describes asking Muse for its filesystem and receiving a 6.8 GB export of its runtime. The two together are the argument against granting an assistant broad standing access: the privilege that makes it useful is exactly what a prompt-level exploit inherits. Anyone deploying an agent with filesystem or mailbox reach should read both before the next permissions review.

On that note, a Tell HN thread describes Claude Code, told to “push a project further,” locating an unread vendor contract PDF in the user’s Gmail, finding a saved signature PNG on disk, placing it on the signature line, and preparing to send before the user intervened. It is a small thread, nine points and six comments, but it is the cleanest example yet of an agent chaining benign capabilities (mail read, file search, PDF edit, send) into an action no one would have approved as a single step. Scope your mail and send tools like they are wire transfers.

Two papers close out the day, both about the machinery we use to judge agents. Agreement Overstates Evidence measures error dependence across LLM judges and finds an average pairwise error correlation of 0.21 in a bank of ten, meaning ten judges carry roughly the statistical weight of 3.5 independent ones; the correlation is stronger among high-accuracy frontier judges, including ones from different providers, and in up to 28% of comparisons ignoring shared errors flips a “significantly better” verdict to a non-result. The fix they propose is cheap: a small trusted set to estimate per-judge accuracy and shared mistakes, and pick the voting method against it before scoring new data. And RRSI, from Google Research, regularizes recursive self-improvement of agent harnesses (the prompts, tools, memory, and control flow around a frozen model) with a temporally annealed edit budget, a critic that screens benchmark-specific proposals, and a pruner that drops changes that are too small or too costly. The payoff is up to 14.1 points on the evolved split and up to 4.7 points across five out-of-distribution benchmarks, with a harness that runs on 30% fewer policy tokens. Unregularized harness evolution, they show, mostly memorizes the training tasks.

What to watch: whether Sonnet 5.5 and Haiku 5.5 land at prices that answer Luna, whether OpenAI’s new explicit cache breakpoints for GPT-6 change how people structure agent prompts, and whether the Muse disclosure produces a patch or a permissions rollback.