Claude Opus 5.5 Price Cut Starts an AI Price War on Every Tier

A month of cheaper models

Anthropic opened the price war with Claude Opus 5.5. It is the company's new flagship and comes in roughly 40% cheaper to run than the model it replaces. The company is selling it as near-top-tier performance at a much lower cost21. Over the following weeks, OpenAI and Google made their own moves. By mid-October, all three major US labs had cut prices or launched discounted models at nearly every tier, from flagship systems down to the small models used for high-volume work.

The pattern is clear. Frontier labs are no longer competing only on benchmark leadership. Capability per dollar is now the main selling point, and a stream of cheaper open-weight models, many of them Chinese, is pushing the labs in that direction.

What Anthropic actually cut

The Opus 5.5 headline needs some unpacking. Anthropic lists the model at $4 per million input tokens and $20 per million output tokens21. Several outlets note that this is only about 20% below Opus 5's $5 and $252325. Anthropic gets to its 40% figure because Opus 5.5 also uses fewer tokens to finish the same work, so typical workloads should cost less overall23. One analysis made the point directly: going from a 20% list cut to a 40% running-cost claim depends on how the model behaves, not on the price sheet25. Dianne Penn, Anthropic's head of product management for research and labs, told CNBC that the company is working on making the model's reasoning and answers more efficient so they use fewer tokens25.

The largest single cut was to cache reads. Their price fell 60%, to $0.20 per million tokens. Anthropic says cache reads make up most of the cost in agentic and coding work23. Opus 5.5 is also reported to produce output more than 30% faster than its predecessor23.

The more important claim is about Anthropic's own lineup. The company says Opus 5.5 matches its more powerful Fable 5.1 model on most tasks. Fable 5.1 costs $10 per million input tokens and $50 per million output tokens21. Coverage stressed the word "most": Anthropic is not promising the same results on every job, so buyers will still need to test their own workloads27. On Terminal-Bench 4.0, an agentic coding benchmark, Anthropic published a score of 66.4% for Opus 5.5. It reported 57.9% for OpenAI's GPT-6 Astra and 55.8% for Fable 5.125.

Anthropic then moved down its lineup. Claude Sonnet 5.5 followed, described as at least 30% faster than Sonnet 5 and up to 30% cheaper29. On October 7 came Claude Haiku 5.5, the third 5.5-generation model in about a month and the first new Haiku in nearly a year30. Haiku's cut was the steepest. For requests under 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, about 90% below Haiku 4.524. Above that size, prices rise fivefold to $0.50 and $2.50. A new tokenizer also counts the same text as roughly 30% more tokens, so Anthropic's own estimate of typical savings is about 75%, not 90%2426. Anthropic also halved Sonnet 5.5 cache-read prices to $0.10 per million tokens and added monthly API credits for Max and Team subscribers30.

OpenAI's counter: a three-tier GPT-6 ladder

OpenAI's prices now run in three steps. GPT-6 Astra, the flagship, lists at $10 per million input tokens and $50 per million output tokens. GPT-6.1 Sol costs $2 and $10, and GPT-6 Luna costs $0.10 and $0.501113. Sol and Luna are about 50% cheaper than the GPT-5.6 models they replace22. GPT-6 Sol fell from GPT-5.6 Sol's $4 and $20 to $2 and $1014.

Reports disagree on when those cuts took effect. One account says Anthropic and OpenAI released their cheaper models within hours of each other on September 2225. Another says OpenAI set off the next round of price cuts on October 7, alongside the global ChatGPT rollout of GPT-614. A pricing tracker covering the same October 7 event concluded that the rollout widened access but did not add a new API rate15. The most likely explanation is that the API prices were set in September and the October rollout put Sol and Luna in front of ChatGPT users: Sol on paid plans and Luna on the Free and Go tiers16. Treating October 7 as a second price cut appears to overstate it.

OpenAI then released GPT-6.1 Sol at the same $2/$10 rate, with cached input cut to $0.10. The company describes it as near-Astra intelligence for a fifth of the price17. OpenAI also compared itself directly with Anthropic. It says GPT-6.1 Sol averages $5.47 per scientific-research task at maximum effort, against $23.21 for Opus 5.517. That figure comes from OpenAI's own benchmark. A separate per-task index puts Opus 5.5 at about $5.98 per task3. The gap shows how much cost-per-task numbers depend on who designs the test.

At the low end, Haiku 5.5 was clearly priced against OpenAI. Its short-request rates match GPT-6 Luna's, including caching24. The two companies differ on where the higher rates begin. Haiku's higher tier starts above 100,000 tokens, while Luna's surcharge starts only above 272,000 input tokens24. For long-document work, OpenAI's small model keeps a real price advantage.

Google's Gemini play: cheap now, double later

Google's response follows a different strategy. Gemini 4 Argon, its new flagship, became broadly available on October 13 after a limited early release1. It launched at an introductory $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input5. The discount is temporary. Google's developer documentation says prices double on December 1, 2026, to $4 and $204.

The timing looks deliberate. During the promotional window, Argon costs half as much per token as Claude Opus 5.5 and one-fifth as much as GPT-6 Astra1. After December 1, Argon and Opus 5.5 cost effectively the same per token6. One analysis described this as starting low to win workloads and raising prices once teams have built on the model8. On capability, one independent index puts Argon at 52.6, compared with 52.7 for GPT-6 Astra and 57.6 for Claude Opus 5.510.

Google has also kept cutting prices on its smaller Flash models. Gemini 3.8 Flash launched on September 2 at an introductory $0.75 per million input tokens and $3.75 per million output tokens. That was Google's third Flash release in 43 days, and those rates double on January 1, 202710. Even so, Gemini's budget tiers now cost more than Luna and Haiku 5.5. Gemini 3.5 Flash-Lite lists at $0.30 and $2.5024.

Why now: open-weight models and an IPO

Coverage gives two main reasons for the timing. The first is competition from open-weight models, which customers can customize and run on their own hardware27. Data from cloud company Vercel suggests these models, many of them from China, are taking market share from US labs21. One comparison lists DeepSeek V4.1 Flash at $0.15 to $0.30 per million input tokens2. That helps explain why Haiku 5.5 does not set the market floor, even after a 90% cut30.

The second is Anthropic's planned stock listing. The company is reportedly preparing an IPO for mid-November26, widely expected before Thanksgiving30. Cheaper models are a way to grow usage before investors look closely at revenue. There is evidence that customer demand was already moving. Payments data from 70,000 companies showed spending on Anthropic's largest and most expensive model levelling off at about 11% of total spending on its tools, as older and cheaper models proved good enough22. Opus 5.5 looks like Anthropic's response: bring near-Fable performance down to a price customers were actually willing to pay.

The timing also stands out for another reason. The launches came after Anthropic chief executive Dario Amodei called on AI companies to slow frontier development to allow more safety testing27. One commentator argued that releasing more capable models at much lower prices does not look like a slowdown25. Anthropic says Opus 5.5 went through alignment testing and external evaluations, and includes safeguards for cybersecurity and biology27. Even so, cheaper capability means wider deployment.

The takeaway

List prices are becoming a weaker guide to what AI actually costs. Anthropic's 40% and 75% claims rely on token efficiency and tokenizer changes. OpenAI's per-task comparisons come from tests it designed. Google's flagship price lasts only seven weeks. Each company is choosing the measure that favors its own models. For buyers, the best guide is the cost of finishing a real task on their own workload. Token prices alone cannot show that, because the number of tokens used, retries and accuracy all affect the final bill24.

Despite the varied pricing, the overall trend is clear. Near-frontier capability now costs about a fifth to two-fifths of the top flagship rate, and the cheapest tiers have dropped to about ten cents per million input tokens. Anthropic made the first cut, OpenAI matched it and Google undercut both for a limited time. The next important date is December 1, when Google's discount ends and Gemini 4 Argon's price doubles.