Cheap AI Models Lead as Claude, GPT-6 and Gemini Cut Prices

The frontier shifts from intelligence to price

For several years, the main race among AI labs was over who had the most capable model. Since early September, the bigger fight has been over price. Axios called it a pivot from building the best and often most dangerous models to building cheaper ones. It argued that falling prices are good news for an industry whose biggest risk is weak demand, because usage of top models is climbing as their costs drop11.

The pace of releases shows it. In a single week in late September, OpenAI, Anthropic, xAI and several Chinese labs all shipped new models. Most of them paired high intelligence with prices that would have looked unrealistic a few months earlier11. Ars Technica described the OpenAI and Anthropic launches the same way: a small step up in capability in exchange for a large cut in cost19.

Claude: the whole lineup gets cheaper

Anthropic changed its pricing more than any other lab. Claude Opus 5.5 arrived on September 22 as the first model in the Claude 5.5 family. It costs $4 per million input tokens and $20 per million output tokens48. Anthropic says it performs about as well as its premium Fable 5.1 on most work while costing about 40% less to run than Opus 55.

That 40% figure needs some unpacking. The list price is only 20% lower than Opus 5. Anthropic gets to its larger number by pointing to cache reads that are 60% cheaper and to the model using fewer tokens per task198. Cache reads now cost $0.20 per million tokens, and Anthropic says they make up most of the cost of agentic and coding work8. Businesses running long coding agents will probably see the full savings. Simpler workloads will see something closer to the 20% list-price cut.

Sonnet 5.5 followed on September 28 as a faster, cheaper partner to Opus 5.5 aimed at well-defined tasks13. Its cache-read price was halved again, to 10 cents per million tokens, when Haiku 5.5 launched3. Haiku 5.5 arrived on October 7 and made the biggest cut of the three. It costs $0.10 per million input tokens and $0.50 per million output tokens, 90% below Haiku 4.5, for prompts up to 100,000 tokens12.

The outlets that covered Haiku 5.5 agree on the main numbers but put the emphasis in different places. VentureBeat and SiliconANGLE pointed out that the new rates match OpenAI's GPT-6 Luna exactly23. Both also noted that prompts over 100,000 tokens cost five times as much, $0.50 input and $2.50 output, which is only a 50% cut from Haiku 4.529. Anthropic says about 90% of Haiku 4.5's requests fell under that limit. That explains why it describes the average saving as roughly 75% rather than 90%1.

GPT-6: a family built in price tiers

OpenAI started at the top and worked down. GPT-6 Astra launched in limited preview on September 3 and reached the public the next day. The rollout came after OpenAI delayed the model to add safeguards following unsanctioned cyberattacks by its agents in July21. Astra costs $10 per million input tokens and $50 per million output tokens, five times as much as the Sol tier2613.

The cheaper models are where OpenAI is competing hardest. GPT-6 Sol and GPT-6 Luna arrived on September 22. OpenAI told Axios they cut costs for top business customers in half compared with the previous versions11. Ars Technica described the capability gains as modest: depending on the benchmark, a few percentage points better than the models they replace, at half the price19. GPT-6.1 Sol followed at DevDay on September 29, priced at $2 input and $10 output, with cached input at $0.1024.

On October 7, GPT-6 became the default model in ChatGPT, alongside a new feature called Intelligent UI that lets answers include charts, forms and small interactive tools. Paid plans run on Sol. Free and Go users get Luna starting October 82325. Putting a model priced at $0.10 per million input tokens behind ChatGPT's free tier shows how far the cost of serving a capable model has come down20.

Gemini: cheap today, more expensive in January

Google cut prices too, but with a time limit. Gemini 3.8 Flash became generally available on September 2. It was Google's third Flash release in 43 days, after 3.6 and 3.7 Flash44. Google's developer documentation lists an introductory price of $0.75 per million input tokens and $3.75 per million output tokens through December 31. Standard rates of $1.50 and $7.50 take effect on January 1, 202749.

That makes Google the only major lab in this cycle that has published a future price increase. One pricing analysis pointed out an odd result: until January, the newest Flash model costs much less than the older Gemini 3.5 Flash, which is still listed at $1.50 input and $9.00 output42.

Google's top model has followed a similar path. Gemini 4 Argon was first restricted to cyber defenders in Google's Fairwind program20. It opened up through the Gemini API and AI Studio on October 1332. Its introductory price is $2 input and $10 output, rising later to $4 and $202044. Google seems to be using discounts to win developers rather than making a permanent price cut. That approach could backfire if competitors keep their low prices into 2027.

Open-weight models are driving the price cuts

Competition from open-weight models, whose weights anyone can download and run, explains much of this. Ars Technica reported that OpenAI and Anthropic are racing to keep up with open-weight models. Enterprises are increasingly using model routers, which send most requests to cheaper models and only the hardest ones to expensive frontier models19. Axios made a similar point: open-weight providers climbing the leaderboards are pushing the top labs to focus on cost11.

DeepSeek is the clearest example. Its V4.1 Flash, released under the MIT license in September, has a 1 million-token context window39. Axios reported that it beat DeepSeek's own previous flagship on several benchmarks, took first place on OpenRouter, and saw usage jump 172% in a week11. Its peak price is $0.30 input and $1.20 output, and off-peak requests cost half that20.

October brought even more open releases. BenchLM's tracker counted 22 confirmed open-weight launches in the month by October 1532. Meta released the weights for Muse Spark Open on October 15 and called it the most capable US open-weight model to date. It has 816 billion total parameters, 42 billion of them active per token, and a 1 million-token context window34. TechCrunch noted that running it locally still takes at least a terabyte of RAM33. Its license is less open than the name suggests: API providers earning more than $20 million a year must sign a separate agreement with Meta35.

Other launches pointed in different directions:

• Mistral Large 4 debuted on October 6 as an API preview, with weights promised by the end of the month35.
• Reflection's Beam, a 501 billion-parameter model, is due to ship under Apache 2.0 later in October36.
• Liquid AI's LFM2.5-24B-A2B uses only 2 billion active parameters per token32.
• Nvidia's Nemotron 3.5 Bolt fits in 24 GB of memory35.

One caution applies to this group. The strongest US open model is not the top open model overall. BenchLM ranks Xiaomi's MiMo-V2.6-Pro first among open-weight models31. Another ranking puts Chinese models GLM-5.3 and Kimi K3 ahead of Muse Spark Open on the Artificial Analysis intelligence index35.

Where the rankings disagree

All the coverage agrees that prices are falling. It disagrees on which model gives the best value. One ranking names GPT-6.1 Sol the best value near the top of the field14. Another finds GPT-6 Luna cheapest when measured by cost per benchmark task, ahead of MiMo-V2.6-Pro and Haiku 5.516.

The results depend on what gets measured. Price per token, price per completed task and which benchmarks are included all produce different winners. Pricing traps also complicate simple comparisons. Grok 4.7 bills the whole request at double the rate once a prompt reaches 200,000 tokens16. Haiku 5.5 charges five times as much above 100,000 tokens2.

What it means

My reading is that "frontier" now has two meanings: the most capable model, and the best capability for the money. The second one matters more for whether the AI boom pays off. Opus 5.5 still sits at or near the top of several capability rankings1316. Even so, what most businesses will notice this fall is that Luna- and Haiku-class models now cost a dime per million input tokens20.

The evidence that cheaper models are expanding demand comes mostly from usage spikes like DeepSeek's, and that data is real but early11. Whether demand grows enough to justify trillions of dollars in infrastructure spending is still unproven. Labs are betting that it will, while open-weight releases keep pushing prices lower.