Anthropic, Meta and Google have landed on the identical price for a frontier model: $4 per million tokens in, $20 out (Google's once a launch discount ends November 30). Anthropic's cheapest model is now priced 90% below its predecessor for most requests. Yet in one independent estimate, a model with a lower sticker price cost more per task than a rival priced two and a half times higher. Here is the fall price war in one chart, and what it means for your monthly bill.
The Prices, Side by Side
AI companies charge per token, a chunk of text a few characters long, priced per million. Input is what you send; output is what the model writes back. For every model in this chart except DeepSeek, output costs exactly five times input, so the input price is a fair way to compare.
Standard list prices. Haiku 5.5's rate covers prompts up to 100,000 tokens; OpenAI's, up to 272,000. DeepSeek's figure is its September launch rate, which SiliconANGLE reports is unchanged.
Three Price Points, Five Companies
$4 in, $20 out. Anthropic launched Claude Opus 5.5 on September 22 at 20% below Opus 5. Meta announced Muse Wonder on October 12, sold through its API at the same rate, and Google's Gemini 4 Argon lands there after its launch offer ends November 30.
$2 in, $10 out. Claude Sonnet 5.5 kept Sonnet 5's price, though Anthropic says it typically "needs far fewer tokens to do the same work." OpenAI says GPT-6.1 Sol delivers "nearly the same level of intelligence" as GPT-6 Astra at one-fifth of Astra's price, TechCrunch reported. Argon sits here until November 30.
Pennies. Anthropic says Claude Haiku 5.5, launched October 7, is "priced 90% lower than Claude Haiku 4.5 for requests up to 100,000 tokens," and about 75% cheaper to run on average. DeepSeek, a Chinese AI lab, updated its V4.1-Flash model with the same low prices and with weights (the trained model itself) free to download under the MIT license.
Why the Sticker Price Can Mislead
Some models think longer and write more to finish the same job. When VentureBeat covered Muse Wonder, it cited Artificial Analysis, an independent testing firm, estimating the cost of an average task on its test suite at $4.52 for Wonder, $5.98 for Opus 5.5 and $3.26 for GPT-6 Astra, "reflecting differences in token consumption across models." Astra charges two and a half times Opus 5.5's per-token rate yet came out cheaper per task. Anthropic also notes that Haiku 5.5 "uses slightly more tokens per task" than Haiku 4.5. Compare cost per finished task, not per token.
What It Means for Your Monthly Bill
Take an illustrative assistant that drafts replies to 1,000 customer emails a month. Assume each reply reads about 2,000 tokens and writes 500 (our estimate for a short email thread). At list prices, that is about $45 a month on GPT-6 Astra, $18 on Opus 5.5, $9 on Sonnet 5.5 or GPT-6.1 Sol, and 45 cents on Haiku 5.5, before the per-task differences above.
For many small businesses, the token bill was already small. The price war matters most for high-volume, simple jobs like tagging tickets or pulling fields from invoices, where a 90% cut on the cheapest tier adds up fast. Back in March, I wrote that AI got 60% cheaper in two years. This fall it kept falling.
What to Do This Month
- Find out which model your tools call and what you paid last month.
- Run 20 real tasks on a cheaper tier and compare quality and cost per finished task.
- Budget for December if you use Gemini 4 Argon: its price doubles after November 30.
- Test before switching. Anthropic warns that code written for Haiku 4.5 can break on Haiku 5.5.
That cost check is often the first step in our AI automation consulting work.
Sources
- Anthropic, "Introducing Claude Opus 5.5" (September 22, 2026): price and the cut from Opus 5. anthropic.com
- Anthropic, "Introducing Claude Sonnet 5.5" (September 28, 2026): price and token efficiency claim. anthropic.com
- OpenAI, "Changelog" (September 29, 2026): GPT-6.1 Sol price. developers.openai.com
- TechCrunch, "OpenAI launches GPT-6.1 Sol, says it nearly matches GPT-6 Astra and costs less" (September 29, 2026): OpenAI's claim and the one-fifth price. techcrunch.com
- OpenAI, "GPT-6 Astra" model page (undated): $10 input and $50 output per million tokens. developers.openai.com
- Anthropic, "Introducing Claude Haiku 5.5" (October 7, 2026): prices, the 90% and 75% figures, tokens per task. anthropic.com
- Anthropic, "Claude Platform release notes" (October 7, 2026): Haiku 4.5 code can break on Haiku 5.5. platform.claude.com
- VentureBeat, "Meta's new Muse Wonder and Muse Wonder Full models achieve frontier performance, but only the weaker one is available via API" (October 12, 2026): Wonder's price and Artificial Analysis cost-per-task estimates. venturebeat.com
- Google Cloud Blog, "Gemini 4 Argon available on Google Cloud" (October 13, 2026): introductory and standard prices. cloud.google.com
- SiliconANGLE, "DeepSeek releases updated V4.1-Flash model it says completes tasks faster" (October 18, 2026): unchanged prices, launch rate, MIT license. siliconangle.com
Paying more for AI than the work needs?
We will look at which models your tools call, test cheaper options on your real tasks, and show you the cost per finished job before you switch anything.