In the span of three days, Sept 28 to 30, Anthropic, OpenAI, and Google each announced a frontier-class model at $2 per million input tokens and $10 per million output tokens. As of yesterday, Oct 13, all three are available for builders to call. One of those prices has an end date.
That shared price gives you a clean reason to reopen your app's unit economics. Pull 20 real tasks from your logs and find out what one completed job costs before you set your paid tier.
The Signal
Anthropic released Claude Sonnet 5.5 on Sept 28 at $2 per million input tokens and $10 per million output tokens. Cache reads cost $0.20 per million, and cache writes cost $2.50 per million. The list price matches Sonnet 5, but Anthropic says the new model costs up to 30% less per task because it uses fewer tokens and generates more than 30% faster.
Anthropic also says Sonnet 5.5 scored two points below Opus 5.5 on GDPval-AA. Opus costs $4 per million input tokens and $20 per million output tokens. Lovable's CTO reported a third fewer tool calls and roughly half the shell runs with Sonnet 5.5. Those are useful claims to test against your own workload. (Anthropic)
OpenAI announced GPT-6.1 Sol on Sept 29 at the same $2/$10 price. Cached input is $0.10 per million, and cache writes are $2.50 per million. OpenAI says Sol approaches GPT-6 Astra on several evals at one-fifth of Astra's standard token price.
Watch the context window before you route long jobs to it. Any prompt over 272K input tokens is billed at twice the input rate and one and a half times the output rate for the whole request. Batch and Flex are 50% off. Sol is available in the API, ChatGPT Work, and Codex. (OpenAI, InfoQ)
Google announced Gemini 4 Argon on Sept 30 and opened it to everyone on Oct 13. It is now on the Gemini API paid tier in AI Studio as gemini-4-argon, and in the Gemini app for Google AI Ultra and Pro subscribers. The launch price is $2/$10, with cached input 95% off.
That price is introductory. It ends Nov 30, 2026, and then Argon moves to $4 per million input tokens and $20 per million output tokens. Caching is implicit only, so you cannot pin an explicit cache the way you can on the other two. Test it now, but set your margin on the $4/$20 rate. (Google, Gemini API docs)
The Stack
Use work your product already performs.
- Pull 20 real tasks from your logs. Keep the input, output, completion result, and model used for every attempt.
- Define what counts as completed before you run the comparison.
- Run the same tasks on
claude-sonnet-5-5,gpt-6.1-sol, andgemini-4-argon. - Calculate pass rate and cost per completed task for each model. Failed attempts still belong in total spend.
- Flag every Sol request over 272K input tokens so its higher rate does not disappear inside an average.
- Price Argon twice: at $2/$10 for today's test, and at $4/$20 for the margin you will actually live with after Nov 30.
If you use Sonnet with thinking turned off, switch to the new between_tools setting before moving to 5.5.
One decision: name the default model for your product's main job, then set the paid tier's price and margin from the measured winner.
Prompt of the Day
Paste this with the task runs from your logs:
Compare Claude Sonnet 5.5, GPT-6.1 Sol, and Gemini 4 Argon using only the rows I provide. Each row may include task name, model, input tokens, output tokens, cached tokens, and whether the task completed.
Use $2 per million input tokens and $10 per million output tokens unless I provide another applicable rate. For every GPT-6.1 Sol request above 272K input tokens, apply twice the input rate and one and a half times the output rate to the whole request. For Gemini 4 Argon, also report total spend at $4 per million input tokens and $20 per million output tokens, the rate after Nov 30, 2026.
For each model, report total spend, completed tasks, pass rate, and cost per completed task. Include failed attempts in total spend. List every Sol request that crossed 272K input tokens.
Finish with a comparison table and identify the lower-cost model among the tasks that passed. If a field is missing, mark it missing. Do not estimate token counts, completion results, cache usage, or prices.
If it fills a missing cell with a guess, throw the comparison out.
Builders in AI Freedom Lab are comparing what one completed task costs before they price an app.
Tool of the Day
Claude, featured in the AI Tools Index.
Sonnet 5.5 earns the slot because lower token use can matter more than matching list prices. Route the 20-task set through claude-sonnet-5-5, record every failed attempt, and compare the completed-task cost with Sol and Argon.
Quick Hits
Anthropic launched Claude Frontier Academy on Oct 2 with $100 million to train 10,000 "Frontier Deployed Engineers" by the end of 2027. The first cohorts come from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley, and Novo Nordisk. Entry is nomination only. Deployment skill is the scarce part. (Anthropic)
Anthropic's submission portal for Claude plugins and MCP connectors is open to developers on Pro, Max, Team, and Enterprise plans. Developers submit to the Claude directory rather than a separate Marketplace submission system. For a small product, that directory is a distribution channel you can ship into. (Tech Help Canada)
Google Cloud AI Research found that agents editing their own harness can overfit to their test tasks. Its regularized method gained up to 14.1 points on trained tasks and up to 4.7 points across five unseen benchmarks while using roughly 30% fewer runtime tokens. Claude Opus 4.8 stayed frozen as the model. Keep a holdout set your agent never tunes on. (The Decoder)
Closing the loop
The $2/$10 tier gives builders a better comparison point. Your logs still decide the winner. Run the same 20 tasks, count failed attempts, check Sol's 272K boundary, price Argon at its post-November rate, and price your app from completed work.
Join AI Freedom Lab, where builders launch companies and web and mobile apps with AI: skool.com/aifreedomlab