{"schema":"sanctity-signal/0.2","record_type":"signal","id":"openai-anthropic-price-cut-e030","domain":"markets","drop_id":"signal-e030","occurred_at":"2026-09-22","how_to_use":"This is a structured record of one real-world action and its competing interpretations, from Sanctity Signal. To build content from it (a reel script, a blog, a thread): open on the hook, ground the viewer with plain and observed, walk the dated trail, present the interpretations as live tension between named readings, and close on the before/after shift and what_to_watch. Hard rules: items under established are fact; items under unknown are never presented as fact; the evidence status (verified or reported) is stated, not dropped; sources are cited.","actor":"OpenAI and Anthropic","action_type":"simultaneous model releases with published price cuts","observed":"Tuesday September 22, 2026. Anthropic published Claude Opus 5.5 at 4 dollars per million input tokens and 20 dollars per million output tokens, against 5 and 25 for Opus 5, with cache reads cut from 50 cents to 20 cents per million. Its stated claim: performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. Published scores: 66.4 percent on Terminal-Bench 4.0 against Opus 5's 52.3 percent, 54.4 percent on FrontierCode, an Elo of 1846 on GDPval-AA v2.1 across 44 occupations, and 81.8 percent partial success on OSWorld 2.0 computer use. Output is generated about 30 percent faster than Opus 5 and subscription usage limits were widened. It ships on Anthropic's own platform and on Amazon Web Services, Google Cloud and Microsoft Azure as claude-opus-5-5, with Sonnet 5.5 and Haiku 5.5 promised within weeks. Three safeguards ship with it: cybersecurity requests route by default to an older model, Claude 4.8, unless the user is enrolled in Anthropic's Cyber Verification Program; biology work is gated behind its Life Sciences Verification Program; and a preserved thinking feature is described as a defence against distillation. The same day OpenAI published GPT-6 Sol and GPT-6 Luna. Sol costs 2 dollars in and 10 dollars out per million tokens up to 272,000 input tokens and 4 and 15 above that, with cached input from 20 cents; Luna costs 10 cents in and 50 cents out. OpenAI describes both as half the price of their GPT-5.6 predecessors. Sol scores 68.8 percent on DeepSWE v1.1, which OpenAI puts at roughly 80 percent cheaper than Claude Fable 5 at that level; Luna scores 66.6 percent on the same test. Sol carries a 1,050,000 token context window and a 128,000 token maximum output. Luna reaches free and Go users in the desktop app; Sol goes to paid ChatGPT Work and Codex plans and to the API. OpenAI's line for the pair: advancing the frontier on cost efficiency. Neither company announced any price change for its top tier: GPT-6 Astra and Anthropic's Fable and Mythos lines were not repriced. The setting: on September 12 Anthropic chief executive Dario Amodei published We Must Pace the Frontier; Sam Altman and Elon Musk backed it publicly the following day. On September 18 four ChatGPT and Claude subscribers filed an antitrust claim in federal court alleging that Anthropic, OpenAI, Google and SpaceXAI met to slow AI development, carded in Edition 027.","corrections":[],"evidence":{"status":"verified","note":null,"sources":[{"role":"primary","label":"Anthropic: Introducing Claude Opus 5.5, Sep 22","url":"https://www.anthropic.com/claude-opus-5-5"},{"role":"primary","label":"OpenAI: Introducing GPT-6 Sol and Luna, Sep 22","url":"https://openai.com/index/introducing-gpt-6-sol-and-luna/"},{"role":"corrob","label":"SiliconANGLE: Anthropic releases Claude Opus 5.5 and OpenAI counters with two cheaper GPT-6 models, Sep 22","url":"https://siliconangle.com/2026/09/22/anthropic-releases-claude-opus-5-5-and-openai-counters-with-two-cheaper-gpt-6-models/"},{"role":"corrob","label":"Fortune: OpenAI and Anthropic release dueling more affordable models as AI price wars heat up, Sep 22","url":"https://fortune.com/2026/09/22/what-ai-slowdown-openai-anthropic-release-dueling-moreaffordable-models-as-ai-price-wars-heat-up/"}]},"strength":{"cost":"Medium","commitment":"High","reversibility":"Medium","persistence":"Months","rationale":"Cost is medium and we are not inflating it. Both companies say the cheaper models are cheaper to serve, so a published price cut of 40 to 50 percent is not necessarily 40 to 50 percent of margin; what it costs for certain is the option to charge more, and the ability to claim later that they were holding back. Commitment is high because API prices are published, are written into enterprise contracts, and are the single number a buyer compares; raising one back is a visible defeat in a way that a withdrawn blog post is not. Reversibility is medium rather than high for that reason, though neither company has promised to hold the price for any period. Persistence is months: this is a competitive position, not a structure, and the next model generation resets it. The reason to weigh it at all is the conjunction, not the amounts."},"shift":{"prior":"Both labs had spent ten days publicly arguing that frontier development should be paced, and four days being sued by their own subscribers for allegedly agreeing to pace it.","action":"Tuesday September 22, 2026. Anthropic published Claude Opus 5.5 at 4 dollars per million input tokens and 20 dollars per million output tokens, against 5 and 25 for Opus 5, with cache reads cut from 50 cents to 20 cents per million. Its stated claim: performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. Published scores: 66.4 percent on Terminal-Bench 4.0 against Opus 5's 52.3 percent, 54.4 percent on FrontierCode, an Elo of 1846 on GDPval-AA v2.1 across 44 occupations, and 81.8 percent partial success on OSWorld 2.0 computer use. Output is generated about 30 percent faster than Opus 5 and subscription usage limits were widened. It ships on Anthropic's own platform and on Amazon Web Services, Google Cloud and Microsoft Azure as claude-opus-5-5, with Sonnet 5.5 and Haiku 5.5 promised within weeks. Three safeguards ship with it: cybersecurity requests route by default to an older model, Claude 4.8, unless the user is enrolled in Anthropic's Cyber Verification Program; biology work is gated behind its Life Sciences Verification Program; and a preserved thinking feature is described as a defence against distillation. The same day OpenAI published GPT-6 Sol and GPT-6 Luna. Sol costs 2 dollars in and 10 dollars out per million tokens up to 272,000 input tokens and 4 and 15 above that, with cached input from 20 cents; Luna costs 10 cents in and 50 cents out. OpenAI describes both as half the price of their GPT-5.6 predecessors. Sol scores 68.8 percent on DeepSWE v1.1, which OpenAI puts at roughly 80 percent cheaper than Claude Fable 5 at that level; Luna scores 66.6 percent on the same test. Sol carries a 1,050,000 token context window and a 128,000 token maximum output. Luna reaches free and Go users in the desktop app; Sol goes to paid ChatGPT Work and Codex plans and to the API. OpenAI's line for the pair: advancing the frontier on cost efficiency. Neither company announced any price change for its top tier: GPT-6 Astra and Anthropic's Fable and Mythos lines were not repriced. The setting: on September 12 Anthropic chief executive Dario Amodei published We Must Pace the Frontier; Sam Altman and Elon Musk backed it publicly the following day. On September 18 four ChatGPT and Claude subscribers filed an antitrust claim in federal court alleging that Anthropic, OpenAI, Google and SpaceXAI met to slow AI development, carded in Edition 027.","update":"Read the pacing argument as a claim about top models only, and watch the tier below it, where price is the weapon, the cuts are 40 to 50 percent, and both labs fired on the same afternoon."},"inference":{"established":["Anthropic released Claude Opus 5.5 on September 22, 2026 at 4 dollars per million input tokens and 20 dollars per million output tokens, against 5 and 25 for Opus 5, with cache reads cut from 50 cents to 20 cents.","OpenAI released GPT-6 Sol at 2 and 10 dollars per million tokens and GPT-6 Luna at 10 and 50 cents on the same day, each described by OpenAI as half the price of its GPT-5.6 predecessor.","Both releases are in the tier below each company's top model, and neither company announced a price change for GPT-6 Astra or for Anthropic's Fable and Mythos lines.","Anthropic shipped Opus 5.5 with cybersecurity requests routed by default to an older model and biology work gated behind an enrolment programme."],"unknown":["That either company abandoned its pacing position. Both argued about the top tier and neither changed the top tier, which is consistent with the argument as stated.","That the price cuts reflect cost. Neither company published a serving cost, and a price can be set below cost to take share.","That the scores are comparable. Each company chose its own benchmarks and its own comparison model, and the two sets do not overlap cleanly.","That this helps or hurts the antitrust plaintiffs. A same-day price war is evidence a court may read as competition or as parallel conduct, and no judge has seen it."]},"narrative":{"hook":"OpenAI and Anthropic cut their prices against each other on the same day, ten days after both endorsed slowing down.","plain":"On September 22 Anthropic released Claude Opus 5.5 at 4 dollars per million input tokens and 20 dollars per million output tokens, down from 5 and 25, and said it costs 40 percent less to run than Opus 5 on typical work while scoring above it. Hours later OpenAI released GPT-6 Sol at 2 and 10 dollars and GPT-6 Luna at 10 and 50 cents, each about half the price of the model it replaces, and undercut Anthropic at every tier. The caveat that matters is that neither company touched the price of its top model: both shipped cheaper models that score close to the expensive ones, which is not the same as making the frontier cheaper.","story":["Look at what each company did not do. Anthropic did not reprice Fable or Mythos. OpenAI did not reprice Astra. The frontier tier, the one both chief executives spent ten days saying should be approached more slowly, is exactly where nothing moved. What moved is the tier underneath, and it moved hard: Opus 5.5 at 40 percent less than Opus 5 while beating it by fourteen points on Terminal-Bench, Sol at half the price of the model it replaces, Luna at ten cents per million tokens in. Anthropic's release also carries the most explicit safety gating it has shipped, routing cybersecurity questions to an older model and putting biology behind an enrolment programme, which is what pacing looks like when it is a product decision rather than an essay.","The timing is what makes this a signal rather than a product note. September 12, Amodei publishes the pacing argument. September 13, Altman and Musk endorse it. September 18, four subscribers sue all four major labs alleging they met to slow development, which is the claim that a pacing agreement is a cartel. September 22, the two named defendants ship cheaper, better models within hours of each other and undercut one another on price. Whatever a court eventually makes of it, four days after being accused of agreeing not to compete, they competed in public on the one variable nobody can misread. Read the pacing argument as a claim about top models only, and watch the tier below it, where price is the weapon, the cuts are 40 to 50 percent, and both labs fired on the same afternoon."],"trail":[{"d":"SEP 12","t":"Dario Amodei publishes We Must Pace the Frontier, arguing for coordinated slowing of frontier development."},{"d":"SEP 13","t":"Sam Altman and Elon Musk publicly back the argument. Microsoft's conduct commitments follow, carded in Edition 021."},{"d":"SEP 18","t":"Four ChatGPT and Claude subscribers sue Anthropic, OpenAI, Google and SpaceXAI for alleged pacing collusion. Carded in Edition 027."},{"d":"SEP 22","t":"Anthropic ships Opus 5.5 at 40 percent lower cost; OpenAI ships GPT-6 Sol and Luna at half the previous tier's price and undercuts it."},{"d":"Reacts","r":true,"t":"The plaintiffs in the Buist antitrust case, whose complaint alleges an agreement not to compete and who now have a same-day price war to quote; the Department of Justice Antitrust Division and the Federal Trade Commission, both of which have open AI matters; Google and SpaceXAI, the other two named defendants, who now choose between matching and being the expensive option; every enterprise buyer who signed an annual contract at the old rates; the cloud resellers, since Opus 5.5 launched simultaneously on Amazon, Google and Microsoft infrastructure; Nvidia and the memory suppliers, because a 50 percent price cut is a bet on inference volume; the AI data suppliers whose contracts are the input cost under these prices; and Congress, where the superintelligence bill and the Senate subcommittee investigation both assume the labs are racing."}],"glossary":[{"term":"Token","def":"The unit AI companies bill by, roughly three quarters of a word. Prices are quoted per million tokens and split between what you send in and what the model writes back."},{"term":"Frontier model","def":"A company's most capable model, the one its public claims rest on. The pacing argument is about these, not the cheaper models sold underneath them."},{"term":"Cache read","def":"The discounted rate for text the model has already processed and is being sent again. It matters most to agents, which resend the same context at every step, and Anthropic cut it by 60 percent."},{"term":"Distillation","def":"Training a cheap model to copy an expensive one's answers. Anthropic's preserved thinking feature is described as a defence against it."}]},"judgment":{"prompt":"Which explanation do you believe?","interpretations":[{"id":"a","label":"Pacing was always about the top tier, and the market moved one floor down.","text":"Neither company repriced its frontier model. Both shipped the tier below at large discounts. That is exactly what a genuine pacing commitment plus ordinary commercial pressure would look like from the outside.","abstain":false},{"id":"b","label":"It is a price war, and the safety language is not driving it.","text":"Enterprise buyers are auditing AI spending, cheaper competitors are taking share, and two companies cutting against each other by 40 to 50 percent within hours is what that pressure produces regardless of what anyone wrote ten days earlier.","abstain":false},{"id":"c","label":"A move timed against each other and against the lawsuit.","text":"Four days after being sued for allegedly agreeing not to compete, the two defendants competed in the most visible way available. The release dates were close enough that at least one of them moved to match.","abstain":false},{"id":"d","label":"The ordinary cadence of model releases.","text":"Point releases arrive every few weeks, cheaper tiers always follow expensive ones, and prices in this industry have fallen steadily for three years. Nothing about September 22 breaks that pattern.","abstain":false},{"id":"e","label":"Too early to call.","text":"The evidence cannot yet separate these.","abstain":true}],"distribution":{"status":"held back until you have made your call","calls":null}},"story_thread":null,"what_to_watch":[{"linked_interpretation":"a","direction":"weakens","observable_event":"Either OpenAI or Anthropic cuts the published price of a top-tier model, GPT-6 Astra or the Fable or Mythos lines, by December 31, 2026.","window_end":"2026-12-31","state":"unresolved"},{"linked_interpretation":"b","direction":"strengthens","observable_event":"A third frontier lab, Google or SpaceXAI among them, cuts published prices by 40 percent or more on a comparable tier, by December 31, 2026.","window_end":"2026-12-31","state":"unresolved"}]}