On September 22, OpenAI shipped GPT-6 Sol and GPT-6 Luna, and Anthropic shipped Claude Opus 5.5. Same day, same pitch: a cheaper, faster version of last quarter's flagship, priced to get you off whatever you were running before. Sol drops from $4 to $2 per million input tokens and $20 to $10 on output, half of what its predecessor cost. Luna goes to $0.10 in and $0.50 out. Anthropic's cut is smaller on paper, 20% off Opus 5's rate to $4/$20 per million tokens, but the cache-read price fell 60%, and Opus 5.5 now performs at roughly the level of Claude Fable 5.1 on most work while running that much cheaper. OpenAI says Sol makes about half as many mistakes as GPT-5.6 Sol on its internal factuality checks, which is a claim worth treating the way you'd treat any vendor's own benchmark, with mild suspicion and a plan to test it on your own workload before you trust it.
What's actually going on here isn't a capability story. Neither company is claiming these are their smartest models; that title still belongs to Astra and Fable. This is both labs racing to cut the price of "good enough for most of what you'll actually ship," and racing there on the same afternoon. A year ago the competitive story was leaderboard position. Now it's cost per token for the tier of model most products are actually built on, and that tier is getting cheaper fast enough that pinning your integration to a specific price point is a mistake. If your margins depend on this month's API pricing, budget for it dropping again before the quarter is out, and don't build anything that only works because the current price happens to be low.
The other thing worth noticing: this is now the second time in three weeks I've written about a same-day release from two competing labs. Astra and its competitors landed close together, and now Sol/Luna and Opus 5.5 did the same thing. Either everyone's on similar training and eval timelines, or nobody wants to be the one who ships into a news cycle the other guy already owns. Both are plausible. Neither makes the "which model should I actually use" decision any easier, since you now have two nearly-simultaneous options claiming the same value proposition and no time to run a real comparison before the next pair shows up.
Meanwhile, VS Code 1.139 shipped on September 23 with a feature that follows directly from something I flagged here after the 1.138 release: agent sessions running inside a project's Dev Container instead of directly on your machine. That containment only worked for local folders before. 1.139 extends it to projects on SSH, Tunnel, and WSL hosts, so an agent working against a remote box gets the same fenced-off environment an agent working locally already had. It's a sensible, unglamorous extension of a good default, and if you're running agents against remote dev boxes you should turn it on rather than assume the container boundary was already there.
The smaller, stranger detail in that same release: VS Code's release notes have carried a disclaimer for a while now, noting that the notes themselves are drafted with GitHub Copilot and might contain inaccuracies. It's gone from the 1.139 page. According to Visual Studio Magazine, Microsoft also went back and quietly removed it from the 1.138 notes, after the fact, with no changelog entry and no comment. Maybe it was never load-bearing and someone decided a standing "this document may be wrong" notice wasn't worth keeping. Maybe it just looked bad next to a release whose whole headline feature is trusting agents with more autonomy. I don't know which, and Microsoft hasn't said. But a documentation team doesn't usually go back and edit a shipped release's notes without a reason, and the reason not being stated is the interesting part.