OpenAI and Anthropic shipped cheaper frontier models 90 minutes apart, and the fight moved to price

Two of the largest AI labs shipped frontier models within 90 minutes of each other on September 22, 2026. Anthropic released Claude Opus 5.5 first; OpenAI followed with GPT-6 Sol and Luna. The timing was not the only thing they shared. Neither company led with a raw capability leap. Both led with price, framing the release around lower cost per token and faster output, and both claimed, in the same afternoon, to have the strongest models available. When two competitors cut prices on the same day and each says it is now ahead, the more useful read is not who won the afternoon but what the pricing says about where the competition has moved.

What Anthropic shipped

Claude Opus 5.5 is available as claude-opus-5-5 across the Claude API, Amazon Web Services, Google Cloud, Microsoft Azure, Claude.ai, and Claude Code. The pricing is the headline: $4 per million input tokens and $20 per million output tokens, each 20 percent below Opus 5’s $5 and $25, with cached reads dropping 60 percent to $0.20 per million. Anthropic puts the blended saving at around 40 percent on typical workloads and says the model runs more than 30 percent faster than Opus 5. On its own agentic-coding benchmarks the jump is large: Terminal-Bench 4.0 rises to 66.4 percent from Opus 5’s 52.3 percent, and CursorBench 4.0 to 57.8 percent from 46.6 percent. Anthropic’s framing is that Opus 5.5 performs at the level of its larger Fable 5.1 model on most work while costing far less, the same move it made with Fable 5.1 a few weeks earlier: push the price down on the workloads that actually spend, the long agent loops that re-read large contexts every turn.

What OpenAI shipped

About 90 minutes later, per TechCrunch’s account, OpenAI launched GPT-6 Sol and Luna, a split lineup rather than a single model. Sol targets complex work such as coding; Luna handles high-volume, clearly scoped tasks like summarizing documents, extracting fields, and answering quick questions. Both are available in ChatGPT Work and Codex for most paid accounts and through the ChatGPT API, with Luna rolling out more broadly to desktop, Free, and Go users. OpenAI prices the 6 series at half the cost of the 5.6 series, crediting improvements in caching and inference, and says Sol makes about half as many mistakes as its predecessor while reaching what it calls Astra-level reliability at much lower cost. The company also asserts both models handle tasks substantially better than Anthropic’s Fable and Opus. As is typical for a launch-day announcement, those figures are OpenAI’s own and had no independent verification at release.

The claims that cancel out

Read side by side, the capability claims mostly neutralize each other. Anthropic calls Opus 5.5 the strongest model it has tested; OpenAI says GPT-6 beats Anthropic’s top models; both cite internal benchmarks that no outside party had reproduced on launch day. That pattern is the thing worth naming, not the individual numbers. The centre of gravity in this release cycle was cost and efficiency, not a new capability frontier, and the framing around it was telling. Anthropic’s own leadership has recently talked about pacing the rate of capability advancement, and this release reads as consistent with that: a cheaper, faster model that matches the prior flagship rather than a clear leap past it. When the marquee number in dueling launches is dollars per million tokens, the frontier is competing on unit economics.

What it means for developers

For a developer picking a model, the practical decision has shifted. A year ago the question was which lab had the clearly stronger model; today, for a large share of coding and agent work, two labs are trading the lead on self-reported benchmarks while both cut prices. That makes cost, latency, and your own evaluations the deciding factors rather than a vendor’s leaderboard. The cache-read economics matter most here: agent loops re-read a large, mostly static context on every turn, so Anthropic’s 60 percent cut to cached reads and OpenAI’s caching-driven price halving both land hardest on exactly the long-running agentic workloads that were most expensive to run. The right response to launch-day benchmark claims is the same as always: run your own tasks on both before you migrate, because the gap between a keynote table and your workload is where switching costs hide.

What to watch

The signal to watch over the next quarter is whether independent evaluations reproduce either lab’s coding and reliability claims outside their own tables, and whether the price cuts hold once real usage, not benchmark runs, meets the bill. The second signal is structural: if the competition is now primarily on cost and latency rather than a capability lead, the labs that win developer default status will be the ones that pair a good-enough frontier model with the cheapest agent loops, not the ones with the highest single benchmark. Same-day launches with dueling price cuts are what a maturing market looks like. Stackmaven’s follow-up will revisit the independent-benchmark picture around December 22.

Sources cited

1. Introducing Claude Opus 5.5 (Anthropic) www.anthropic.com
2. Introducing GPT-6 Sol and Luna (OpenAI) openai.com
3. Anthropic releases Opus 5.5 with lower prices and Fable-level performance (TechCrunch) techcrunch.com
4. OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes (TechCrunch) techcrunch.com