Anthropic and OpenAI Cut AI Prices by Half Days After Their CEOs Agreed to Slow Down

Introduction

On the afternoon of Tuesday, 22 September 2026, two of the most direct rivals in artificial intelligence shipped cheaper models within hours of each other. Anthropic released Claude Opus 5.5, its first entry in a new model family, priced at $4 per million input tokens and $20 per million output tokens. OpenAI answered with GPT-6 Sol and GPT-6 Luna, two tiers beneath its GPT-6 Astra flagship, with API rates cut by roughly half. OpenAI also confirmed to VentureBeat that the new rates are permanent list prices rather than promotional introductory pricing.

The timing is the story. Ten days earlier, on 12 September, Anthropic chief executive Dario Amodei published an essay urging the industry to "pace the frontier" and slow the development of the most capable systems. Sam Altman of OpenAI, Elon Musk of SpaceXAI and Demis Hassabis of Google DeepMind each publicly endorsed the proposal the same day. Six days after that, on Friday 18 September, paid subscribers to ChatGPT, Claude, Grok and Gemini filed an antitrust lawsuit in the US District Court for the Northern District of California alleging the four companies had made an illegal pact to slow AI progress.

Then the two largest labs cut prices in half. Fortune reported the releases as the opening move of a price war that "could put pressure on their ability to profit from their models down the line." The Financial Times framed it as competition "between the two leading US labs." What neither headline confronted is that a genuine slowdown and a genuine price war pull in opposite directions.

The Price Table

The concrete numbers matter more than the framing. Claude Opus 5.5 is priced at $4 per million input tokens and $20 per million output, against $5 and $25 for Opus 5, released in July. Anthropic says the model delivers performance comparable to its more capable Fable 5.1 model, which costs $10 and $50, on most work. Sonnet 5.5 and Haiku 5.5 are expected "over the coming weeks."

The largest single cut is in caching. Opus 5.5 cache reads fall from $0.50 to $0.20 — 40% of the old price — while cache writes fall from $6.25 to $5. Anthropic added a fast mode at $8 and $40 for roughly 2.5x the speed. For any workload with a large stable prefix, which describes most agentic and document processing jobs, the cache cut is worth more than the headline rate reduction.

OpenAI's tiers are cheaper still. GPT-6 Sol lists at $2 per million input tokens and $10 per million output, down from $4 and $20 for GPT-5.6 Sol. GPT-6 Luna lists at $0.10 and $0.50, down from $0.20 and $1.20 — an output reduction of roughly 58%, deeper than the headline 50%. Luna is the free tier for ChatGPT users, a distribution decision as much as a pricing one.

Run a representative job — one million input tokens, 100,000 output tokens — and the spread becomes stark. Opus 5.5 costs $6.00. GPT-6 Sol costs $3.00. GPT-6 Luna costs $0.15. On that shape of work, Luna is one fortieth the cost of Opus 5.5.

Both labs attribute the reductions to efficiency rather than subsidy. OpenAI credits "improvements in caching and inference." Dianne Penn, Anthropic's head of product management, research and labs, described work on "how to make that thinking, how to make the answering more efficient, so it uses less tokens depending on your effort setting." Cheaper per token plus fewer tokens is how a 20% list cut becomes a 40% running-cost claim.

What The Benchmarks Actually Show

Anthropic says Opus 5.5 posts 66.4% on Terminal-Bench 4.0, an agentic coding benchmark, against 57.9% for GPT-6 Astra, 55.8% for Fable 5.1 and 52.3% for Opus 5. It leads FrontierCode v1.1 at 54.4% and Humanity's Last Exam at 67.7%.

Customer-reported results are more revealing than the leaderboards. Deloitte reports Opus 5.5 catching 72% of bugs in code review at a low effort setting, against 56% for Opus 5 running at high effort. Hebbia reports 86.6% coverage on end-to-end finance workflows against 60.3%. Box reports answers 40% less verbose without losing accuracy, and Optiver reports matching Opus 5 quality in roughly half the turns and output tokens. These are bill reductions that never appear on a price page, and they are the reason the efficiency claim is credible even though a pure list-price comparison yields only 20%.

The safety picture moved too. Anthropic reports Opus 5.5 attempted to circumvent boundaries about 85% less often than Opus 5 across a behavioural audit of nearly 2,000 scenarios, and it ties Fable 5.1 for the lowest attack success rate on Gray Swan's prompt injection evaluation. The model ships with the same tier of cybersecurity, biology and frontier-development safeguards previously reserved for Fable 5.1 — a meaningful concession, since Anthropic had described Opus 5.5 as undergoing its most rigorous alignment testing to date with outside evaluation groups.

The Open-Weight Floor

Neither lab mentions it in the release notes, but the more durable pressure comes from outside. CNBC reported that both companies face "stiff competition from cheaper open-weight models," naming Alibaba, Moonshot AI and DeepSeek among the firms applying downward pressure on US pricing.

An open-weight model you can host yourself sets a hard ceiling on what a proprietary API can charge for commodity work. Extraction and summarisation are commodity work. GPT-6 Luna at $0.10 per million input tokens is precisely a response to that ceiling — and it simultaneously raises the bar for the open-weight hosts that started the pressure, since self-hosting now has to beat $0.10 per million tokens on total cost including operations.

Why The CFO Is Buying

The customer-side driver is measurable. Randall Hunt, chief technology officer at the AI consultancy Caylent, told Fortune that "CFOs have seen some of the sticker shock, and they haven't seen some of the gains that were promised in the initial investments." His firm now measures cost per completed task rather than cost per thousand tokens.

Ara Kharazian, lead economist at Ramp, offered the sharper warning: "AI bulls assume that there will be highly performant models that provide more and more value, and therefore they should be more expensive. But that is not how normal technology makes it to market." His point is that a price war is a competitive state, not a trend, and neither lab is obliged to keep cutting.

Anthropic is also preparing for a potential initial public offering, which makes the margin question existential rather than theoretical. Investors are watching whether the company can sustain rapid growth as competition intensifies and buyers become more cost-sensitive. That pressure is part of what makes the simultaneous release rational — and part of what makes the earlier slowdown essay look, in retrospect, less like a constraint on the business than a positioning exercise in public debate.

The Antitrust Shadow

The lawsuit filed in San Francisco alleges the coordination "largely took place on Sept 12," when Amodei's essay appeared and three rival chief executives endorsed it in public. The plaintiffs, represented by lead attorney Nick Rowley, argue it is clear that an agreement among AI's chief rivals that progress "should be slower than competition would otherwise produce has an anti-competitive effect on consumers." Their theory is that slower capability development reduces what subscribers receive for the money they pay.

Amodei himself acknowledged the antitrust exposure, writing that it would help for the US government to mediate "or at least enable" cross-lab discussions, ideally issuing "a narrow waiver for certain kinds of safety conversations." Altman responded that OpenAI welcomed a federal framework setting consistent safety requirements but did not "need to wait for an anti-trust exemption or legislation to begin the work."

That waiver is not going to arrive easily. Senator Josh Hawley said in a recent Senate hearing that "there is no world" in which he would agree to give "the most powerful companies in the history of the world" an exemption from antitrust laws to collaborate. President Trump rejected calls for regulation on social media, called such efforts a "conspiracy," and said on 19 September he was forming an AI task force and would appoint an "AI czar." Global AI equities fell on 14 September after the industry chief executives called for slowing development.

Conclusion

The two narratives sit uncomfortably together. If the frontier is genuinely slowing, then the price war is the frontier's downstream consequence: labs optimising the capability they already have, squeezing inference cost, and competing for the high-volume, everyday workloads that actually generate revenue. That reading is defensible. Neither Opus 5.5 nor Sol and Luna claims a generation-level capability jump, and a lab can consistently argue it slowed frontier work while making existing capability cheaper.

The harder reading is that price is now the main lever on deployment. Halving the cost of agentic coding does more to increase real-world AI activity over the next quarter than most capability improvements would. Efficiency work is not neutral with respect to how much AI gets used, which is what turns a commercial decision into a safety question. The companies now facing a lawsuit for agreeing to slow down are simultaneously making their existing models dramatically cheaper to deploy at scale.

For buyers, the practical advice is unglamorous. Re-run model selection against a cheaper tier, route traffic by how much judgement a task actually requires, and build cost models on today's list prices rather than an extrapolated curve — Sol and Luna are permanent prices, but nothing obliges either lab to keep cutting. The rest of this coverage of model economics sits under cloud and edge computing.

Images

References

• Fortune, "What AI slowdown? OpenAI, Anthropic release dueling models as price wars heat up," 22 September 2026.
• CNBC, "Anthropic and OpenAI launch cheaper models," 22 September 2026.
• Business Standard (syndicating AP), "Anthropic and OpenAI roll out cheaper models amid intensifying AI race," 23 September 2026.
• Business Standard (syndicating AP), "Anthropic, OpenAI, Google, SpaceXAI face lawsuit over 'AI slowdown pact'," 19 September 2026.
• India Today, "US lawsuit says OpenAI, Anthropic and Google colluded to slow AI development," 19 September 2026.
• Reuters, "Global AI stocks fall as industry chiefs call for slowing development," 14 September 2026.
• Anthropic product documentation and public model cards for the Claude 5.5 family.
• OpenAI API pricing documentation for the GPT-6 Sol and Luna tiers.
• Gray Swan prompt injection evaluation and Anthropic behavioural audit results, as published by Anthropic.