Anthropic and OpenAI Cut Model Prices in Their First Releases Since the Slowdown Debate

Anthropic and OpenAI each released a cheaper AI model on Tuesday, 22 September 2026, the first launches from either lab since the argument over slowing down frontier AI became the industry's loudest conversation. OpenAI added two lower-cost models to its GPT-6 line, Sol and Luna. Anthropic introduced Claude Opus 5.5, the first entry in a new Claude 5.5 family. CNBC reported both releases the same day.

Related reading: Anthropic's Claude Duet Works in the Background, and Asks Before It Acts · Anthropic Cuts Live Internet Access for Its Internal Evaluations After Its Agents Broke the Rules.

The timing carries as much weight as the products. Anthropic chief executive Dario Amodei spent the preceding fortnight arguing that labs should deliberately pace how fast they improve their most capable systems, and his company has now shipped a model whose central selling point is that it is 40 percent cheaper to run than the one it replaces. OpenAI, whose chief executive Sam Altman endorsed that call, halved the price of its mid-tier models. Cheaper inference is not the same thing as a slower frontier, yet the two announcements together show where commercial competition has moved: toward the cost of finishing a task rather than raw capability at any price.

What OpenAI shipped

OpenAI's new models extend the GPT-6 generation that began with Astra, the flagship the company launched earlier in September. Sol sits one tier below Astra and is aimed at complex work such as coding. Luna handles what OpenAI calls high-volume tasks, naming document summarization, information extraction and quick answers. TechCrunch reported that OpenAI describes the pair as extending Astra's intelligence "by making that intelligence more efficient and accessible."

The price list is the clearest signal. OpenAI's own pricing page lists GPT-6 Sol at $2 per million input tokens and $10 per million output tokens, and GPT-6 Luna at $0.10 per million input and $0.50 per million output. The previous generation, GPT-5.6, listed Sol at $4 and $20 and Luna at $0.20 and $1.20, so the new models land at roughly half the earlier rates, matching OpenAI's claim of a 50 percent cut against GPT-5.6 promotional pricing. Cached input is cheaper again, at $0.20 per million tokens for Sol and $0.01 for Luna. Reuters reported the same headline numbers.

There are volume discounts on top. OpenAI's batch tier halves the standard rates to $1 per million input and $5 per million output for Sol, and $0.05 and $0.25 for Luna. Prompts that run past the standard context window pay double, at $4 and $15 for Sol and $0.20 and $0.75 for Luna. A faster serving tier, called Fast mode, doubles the standard price again.

OpenAI credits better prompt caching and inference efficiency for the reduction. ZDNet noted the company's claim that a cached prompt can cost up to 90 percent less than an uncached one, which matters most for long-running agents that reuse the same instructions across many calls. OpenAI also says GPT-6 Sol makes about half as many mistakes as GPT-5.6 Sol on its internal factuality evaluation, reaching Astra-level reliability at lower cost, and that Luna at higher effort levels matches GPT-5.6 Sol at a small fraction of the price. Those are company figures rather than independent measurements.

Availability starts where developers already work. TechCrunch and The Verge reported that both models reached ChatGPT Work, Codex and the ChatGPT API for most paid accounts, with Luna also in the desktop app and open to Free and Go users. OpenAI said it would roll the models out across the ChatGPT app and website through the day.

What Anthropic shipped

Anthropic's Claude Opus 5.5 is the first model in the company's Claude 5.5 family. In its announcement, Anthropic says the model "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5." The pricing reflects part of that: input tokens fall from $5 to $4 per million and output from $25 to $20, a 20 percent cut per token, while cache reads drop from $0.50 to $0.20, a 60 percent reduction. Anthropic says Opus 5.5 also generates output more than 30 percent faster than Opus 5.

The 40 percent figure combines lower per-token prices with fewer tokens used per task, and Anthropic is explicit that it is a test result at default settings on typical workloads rather than a guarantee. The model is available now on Amazon Web Services, Google Cloud and Microsoft Azure, and through the Claude API under the identifier claude-opus-5-5. A fast mode costs $8 per million input tokens and $40 per million output with up to 2.5 times the speed. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks.

The pitch is efficiency as much as price. Dianne Penn, Anthropic's head of product management for research and labs, told CNBC that the company keeps working on how a model spends its reasoning budget.

"One of the things we're continuing to innovate on is how to make that thinking, how to make the answering more efficient, so it uses less tokens depending on your effort setting." — Dianne Penn, head of product management, research and labs at Anthropic

Anthropic also frames Opus 5.5 as its first release since Amodei called for pacing the frontier. The company says the model was tested before release by outside evaluators including METR and Frontier Design, and that it carries the same class of safeguards as Fable 5.1 in cybersecurity, biology and distillation, the practice of extracting a model's behavior at industrial scale. In its announcement, Anthropic points to a code migration of 680,000 lines that one early tester finished in under a day, and to an internal test in which Opus 5.5 rewrote the HAProxy load balancer from C into Rust in 9.5 hours, against 12 hours for Fable 5.1. Both are company accounts.

The slowdown backdrop

The context for the launches is an unusually public argument about speed. Amodei published a post titled We Must Pace the Frontier on 12 September, arguing that companies should slow the rate at which they improve model capabilities so that safety work can keep up. CNBC reported that Altman and Elon Musk, the chief executive of Tesla and SpaceX, both endorsed the call. The debate sharpened on 8 September, when Jacob Coxon, a former Anthropic researcher, said on X that he had quit, warning that both labs were "gambling with our lives."

A cheaper model is not a weaker frontier model, and the two ideas can coexist. What the releases show is that the labs' commercial competition now concentrates on cost per task, where buyers compare what it takes to finish a job rather than what a leaderboard reports. Anthropic's own announcement acknowledges the shift, noting that at current capability levels "benchmark margins have become a less reliable guide to real-world differences."

Why cheaper, why now

The immediate pressure is economic and it arrives from several directions. Both labs face cheaper open-weight models from Chinese companies including Alibaba, Moonshot AI and DeepSeek, which customers can run on their own hardware without paying a frontier premium. Enterprises, meanwhile, have been trying to rein in AI spending after two years of rapid adoption. CNBC reported that satisfying cost-conscious customers was a central motive for both releases.

There is a technical reason that cost has fallen as well. Prompt caching, which stores the processed form of a repeated prompt, and more efficient serving mean the same hardware can complete more work. Anthropic says Opus 5.5 requires less compute to serve than Opus 5. OpenAI credits caching and inference improvements for its cuts. When the same accelerator serves more tokens, the price per token can fall without a matching drop in capability.

The competitive geometry matters too. TechCrunch noted that Anthropic released Opus 5.5 about 90 minutes before OpenAI's launch, a gap that says as much about rivalry as the models do. Opus 5 arrived on 24 July, so the successor came roughly two months later. The two companies chase the same enterprise budgets, and each chose price as its headline rather than a new benchmark record.

Vendor benchmarks and their limits

Each lab published benchmarks that favor its own model, and readers should treat them as marketing until independent testing appears. Anthropic's table puts Opus 5.5 at 66.4 percent on Terminal-Bench 4.0, ahead of Fable 5.1 at 55.8 percent and GPT-6 Astra at 57.9 percent. On GDPval-AA v2.1, a test of real-world professional work across 44 occupations, Anthropic reports 1846 Elo for Opus 5.5 against 1735 for Fable 5.1. OpenAI counters with cost-per-task comparisons, including an OSWorld result where it says GPT-6 Sol at maximum effort scores similarly to Claude Opus 5 at medium effort. Both sets are self-reported.

The firmer comparison is the published price list, which is contractual rather than promotional. On that measure, OpenAI's Sol undercuts Anthropic's Opus 5.5 on both input and output tokens by a wide margin, while Anthropic positions its model as a direct replacement for the older Opus 5 at a lower price. Luna is cheaper still, though OpenAI describes it as suited to narrower tasks with a clear goal. ZDNet also made a point that applies to both launches: API price cuts do not automatically reach subscription plans, and neither lab said whether a given task would consume plan usage more slowly.

What to watch

Three questions will decide whether these cuts hold. The first is whether the low prices survive. OpenAI's own page notes that GPT-5.6 Sol's promotional pricing runs at least through 21 November 2026, a reminder that launch rates and list rates differ. Buyers planning budgets around $2 and $0.10 per million tokens should check the terms again in the new year.

The second is whether cheaper models change what teams automate. When a task costs a tenth of what it did, work that was uneconomic to hand to a model becomes worth doing, and that can expand usage faster than the price fell. The efficiency claims from both labs point in that direction, and the coming Sonnet 5.5 and Haiku 5.5 releases will test it at lower tiers still.

The third is whether the gap between the cheap and flagship tiers keeps narrowing. If Luna and Opus 5.5 handle more of the work that once required Astra or Fable 5.1, the premium tier's own price will face pressure. For now, the two labs have made the same bet: in a market worried about both safety and spending, the next round of competition will be decided on the cost of a finished task.

Cover: Wikimedia Commons photo "Dario Amodei at TechCrunch Disrupt 2023 01.jpg" by TechCrunch, licensed under CC BY 2.0.