OpenAI and Anthropic both halved model prices on the same day

Two price cuts, one Tuesday

Anthropic shipped Claude Opus 5.5 on September 22, 2026, and a few hours later OpenAI launched GPT-6 Sol and Luna. Both companies framed the releases around the same idea: same or better quality, meaningfully lower cost per token. That is not a coincidence you can shrug off — it is the first time the two leading frontier labs have cut prices within the same news cycle, and Thurrott's reporting noted Anthropic "changed the competitive picture again Tuesday morning" only for OpenAI to answer with cheaper, faster models before the day was out.

For anyone building a coding agent, a RAG pipeline, or anything else that pays per token, this matters more than the usual model-launch churn. Two products you might already be running in production got cheaper and faster within hours of each other, and the benchmark claims each company is making don't compare apples to apples yet. Before you touch a config file, it's worth working through what actually changed.

The new price-per-token math

Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, down from Opus 5's $5/$25 — a 20% cut on Anthropic's list pricing, which the company frames as roughly 40% cheaper to run in practice once you account for faster generation. Anthropic says Opus 5.5 performs at a level similar to Claude Fable 5.1 on most tasks, generates output about 30% faster than Opus 5, and scored as "the strongest-performing model we've tested to date" on its own automated misalignment audit, with pre-release testing from external evaluators Frontier Design and METR.

GPT-6 Sol, meanwhile, is priced at $2 per million input tokens and $10 per million output — half of what GPT-5.6 cost — and OpenAI's cheaper Luna model comes in at $0.10/$0.50, also a 50% cut from its predecessor. At list pricing, that puts Sol at roughly half the cost of Opus 5.5 on both input and output rates. OpenAI says GPT-6 Sol "makes about half as many mistakes as its predecessor, reaching Astra-level reliability at much lower cost" — a strong claim, but one measured against Sol's own predecessor, not against Opus 5.5.

That last point is the one to sit with. Anthropic released Opus 5.5 hours before OpenAI's scheduled Sol/Luna announcement, which means any benchmark comparison OpenAI has published references Opus 5, the older model, not the one that just shipped. Thurrott flagged this gap explicitly in its coverage. In practice, nobody outside these two companies has run Opus 5.5 and GPT-6 Sol through the same eval suite yet. The headline "half the price, similar quality" is true for each model against its own predecessor — it is not yet a verified claim about one model against the other.

What "cheaper and faster" actually buys a team

If you're choosing between Sol and Opus 5.5 for a coding agent or a high-volume pipeline, the price difference is real and worth planning around, but it isn't the only variable. Sol's "Astra-level reliability at lower cost" is OpenAI's own framing for fewer mistakes relative to GPT-5.6, not a claim that Sol beats Opus 5.5 head to head. Opus 5.5's "40% cheaper to run" bakes in the 30% speed-up, which matters for latency-sensitive agent loops as much as for the invoice at the end of the month — a model that finishes a task in fewer wall-clock seconds is worth more than its raw per-token rate suggests, especially in interactive coding tools.

There's also a distribution angle worth noting. Sol and Luna are rolling out across ChatGPT Work, Codex, and the ChatGPT API for paid accounts, with Luna additionally reaching the desktop app and free and Go tier users — so a large chunk of ChatGPT's existing user base gets the cheaper model by default, not by opt-in migration. Opus 5.5 is live across AWS, Google Cloud, and Microsoft Azure, and Anthropic says this is the first release in a "5.5 wave," with Sonnet 5.5 and Haiku 5.5 due in the following weeks — meaning the pricing picture at the mid and low end of Anthropic's lineup is still incomplete. GitHub's Mario Rodriguez, quoted in Anthropic's own release notes, framed the update around agentic coding specifically: developers want agents that can take on real software work and finish it, and the efficiency gains are being pitched in terms of tokens and steps per completed task, not just dollars per token.

Why waiting a few days is the sane move

None of this means you should switch production traffic today. The two companies are making claims about their own new models against their own old models — a defensible thing to publish, but not the comparison you actually need if you're deciding between Opus 5.5 and Sol for the same workload. Independent, apples-to-apples benchmarks that put both models through the same tasks haven't landed yet, and given that Opus 5.5 shipped only hours before Sol and Luna, even OpenAI's own comparisons couldn't have included it.

The practical move for a team running either model in production: hold your current model choice for a few more days and let third-party evals catch up. When you do test, run your own workload-specific eval — your agent's actual tool-calling pattern, your actual prompt lengths, your actual failure modes — rather than trusting either company's aggregate benchmark numbers. Price cuts of this size are real savings, but only if the model still clears your bar on correctness; a 50% cheaper model that needs a retry loop to match your current error rate isn't actually cheaper.

If you're currently on Opus 5 or GPT-5.6, the cost math alone justifies testing the new versions in a staging environment now — the price drop lands immediately whether or not the quality claims hold up. If you're mid-migration or building a new pipeline from scratch, it's worth waiting for the first independent head-to-head numbers, which given the pace of coverage from outlets like VentureBeat and Thurrott, should surface within the week. Either way, keep an eye on Anthropic's promised Sonnet 5.5 and Haiku 5.5 releases — for most teams running high-volume, latency-sensitive workloads, the mid-tier models, not the flagship Opus and Sol, are where the actual production decision gets made.