Opus 5.5 and GPT-6 Sol/Luna cut workhorse prices the same day. For a small team, the file is routing and unit economics, not brand loyalty.

The model story for Tuesday, September 22, 2026 is not a freeze. It is two cheaper workhorse tiers landing on the same calendar day after both CEOs had just argued for pacing the frontier. Anthropic shipped Claude Opus 5.5 first. OpenAI shipped GPT-6 Sol and GPT-6 Luna about ninety minutes later, per TechCrunch. Today is Wednesday, September 23. Treat Tuesday as the event date and today as the operator follow-through. Do not write that either launch happened today.

I also want the adjacent September posts kept on their own pages. Plugin4Shell is Tuesday's distribution-layer note. Buist is Monday's antitrust filing. Google opening Opus 5 inside Antigravity is September 15. Amodei's pacing essay and the evaluator post sit around September 11 to 13. GPT-6 Astra's computer-use launch is September 4. This page is only the Tuesday price-and-efficiency drop for Opus 5.5 and GPT-6 Sol/Luna.

What each lab claimed on Tuesday

Anthropic's own post dates Opus 5.5 to September 22. It is the first model in the Claude 5.5 family. Anthropic says it performs at Claude Fable 5.1 level on most work and costs about 40% less to run than Opus 5 on typical workloads. List prices are $4 per million input tokens and $20 per million output tokens. Cache reads are $0.20 per million, which matters for agentic and coding sessions where cache traffic dominates the bill. Anthropic also says Opus 5.5 is more than 30% faster on output than Opus 5, and that it was externally evaluated before release (Frontier Design and METR named). Safeguards sit near the Fable 5.1 class for biology and cybersecurity. Those are Anthropic's claims. Cite them as such.

OpenAI's GPT-6 Sol and Luna community announce, also September 22, frames the pair as faster and cheaper siblings that carry Astra-era strengths into scale work. Sol is positioned for complex coding and agentic jobs. Luna is positioned for high-volume clerical work with a clear goal: summarize, extract, answer quick questions. OpenAI says API prices for Sol and Luna are 50% lower than GPT-5.6 Sol/Luna promotional pricing, and attributes the cut to caching and inference efficiency. TechCrunch notes OpenAI also claims GPT-6 Sol makes about half as many factuality mistakes as its predecessor on an internal eval of de-identified conversations where users flagged errors. Competitive jabbing against Anthropic is in the packaging. Do not invent a clean bench sweep either way. Use each lab's own wording and keep Astra's earlier full launch separate from this Sol/Luna refresh.

Pace talk did not freeze cheaper shipping

The tension is the useful part. Amodei's "We Must Pace the Frontier" essay and Altman's public backing of pacing sat in mid-September. Buist filed the slowdown antitrust complaint on September 18, with Monday's reporting still hot. Neither lab answered that week by pausing mid-tier capability. They answered by shipping denser, cheaper workhorse models on the same Tuesday. Pace talk is a governance posture. It is not a freeze on list prices or cache economics. Write that distinction into the vendor file before a headline collapses the two.

Also keep the Sol/Luna claim narrow. OpenAI is not saying Sol replaces Astra for every computer-use job, and Anthropic is not saying Opus 5.5 beats every named competitor on every bench. The operator-relevant move is the middle of the stack getting cheaper and denser for coding agents and high-volume clerical automation at once.

The small-team file

You do not need a brand loyalty vote. You need a short routing and unit-economics pass.

1. Re-price your top three coding-agent workloads. Pull last month's token mix for the jobs that actually burn money: multi-file refactors, overnight agents, long review loops. Apply Opus 5.5's $4/$20 list and $0.20 cache reads, and Sol's new promotional-relative API cut, on the same traces. Keep effort settings and tool-call counts honest. A model that uses fewer turns can beat a cheaper sticker, and the reverse is also true.

2. Re-price your high-volume clerical lane separately. Summarize, extract, classify, and FAQ answering belong on Luna-class economics, not on flagship coding defaults. If you have been routing volume work through a coding-tier model by habit, Tuesday is the day to stop.

3. Watch cache-hit rates, not only list prices. Anthropic's post is explicit that cache reads dominate agentic and coding cost. If your harness does not report cache hits, add that metric before you declare a winner. A 40% run-cost claim evaporates when every session cold-starts.

4. Prefer workload fit over yesterday's flagship name. Sol and Opus-class for complex coding and agentic work. Luna-class for clear-goal volume. Astra stays the earlier computer-use story from September 4. Do not collapse those lanes into one "we use the newest logo" ticket.

5. Keep pacing and antitrust adjacent, not merged. Buist, Altman's mid-September pacing comments, and Amodei's essay are governance context. They do not rewrite Tuesday's price sheet. Dual-path vendor assumptions still matter if discovery or settlement pressure later reshapes calendars. Price the workloads you run today on the prices that shipped Tuesday.

6. Do not wait for a perfect third-party scoreboard. Each lab's benches are competitive packaging. Your merge rate, rework rate, and cost per closed ticket are the scoreboard that pays rent.

If you want a calm second set of eyes on how this lands in a small-team vendor and routing file, DefendResolutions is built for that kind of operator work.

What not to write

Do not write that Opus 5.5 or GPT-6 Sol/Luna launched today. Events were Tuesday, September 22. Do not collapse Astra's September 4 launch with Sol/Luna. Do not fold this into Plugin4Shell, Buist, Antigravity, or the mid-September pacing essays. Do not invent that Sol/Luna beat Opus 5.5 on every bench, or the reverse. The homework is routing, cache-aware unit economics, and workload fit.