In September 2026, the AI labs started a price war. OpenAI launched GPT-6 Sol at $2 per million input tokens on September 22, alongside the budget GPT-6 Luna at $0.10. Anthropic answered with Claude Opus 5.5 the same day and Claude Sonnet 5.5 a week later. OpenAI refreshed its mid-tier with GPT-6.1 Sol at DevDay on September 29. Then on October 1, Google unveiled Gemini 4 Argon at an introductory $2 and $10.
The result is a frontier price list that has been rewritten in two weeks. Here is what a million tokens actually costs right now, and why the number on the price list is not the number on your invoice.
The full frontier price list, October 2026
All figures below are list prices per million tokens in US dollars, with Indian rupee equivalents at roughly Rs 96 to the dollar:
| Model | Input | Output | Input (INR) | Output (INR) |
|---|---|---|---|---|
| GPT-6 Luna (OpenAI) | $0.10 | $0.50 | ~Rs 9.60 | ~Rs 48 |
| GPT-6.1 Sol (OpenAI) | $2 | $10 | ~Rs 192 | ~Rs 960 |
| Claude Sonnet 5.5 (Anthropic) | $2 | $10 | ~Rs 192 | ~Rs 960 |
| Gemini 4 Argon (Google, introductory) | $2 | $10 | ~Rs 192 | ~Rs 960 |
| Claude Opus 5.5 (Anthropic) | $4 | $20 | ~Rs 384 | ~Rs 1,920 |
| Gemini 4 Argon (after intro) | $4 | $20 | ~Rs 384 | ~Rs 1,920 |
| GPT-6 Astra (OpenAI flagship) | ~$10 | ~$50 | ~Rs 960 | ~Rs 4,800 |
Three observations jump out. First, three different labs converged on exactly $2 and $10 within nine days. GPT-6.1 Sol, Claude Sonnet 5.5, and Gemini 4 Argon all landed on the same price pair, which tells you these companies are pricing against each other in real time. Second, the gap between the mid-tier and the flagship is enormous: Astra costs roughly five times what Sol costs per token, per OpenAI’s own “one-fifth” framing. Third, Google’s Argon price is explicitly temporary. Google says the introductory $2 and $10 rate will rise to $4 and $20, landing on par with Claude Opus 5.5, and has given no date for when that happens. Anyone costing a project at Argon launch rates is pricing a window, not a rate.
Cost per token is not cost per task
The price list hides the variable that matters most: how many tokens each model burns to finish a job. On OpenAI’s reported Terminal-Bench Science results, GPT-6.1 Sol averaged $5.47 per completed task, against $23.21 for Claude Opus 5.5 and $23.80 for GPT-6 Astra. Astra still holds the highest score of the group, so the flagship keeps its crown, but the per-task cost gap is roughly four times while the score gap is modest.
Anthropic makes the same argument for Sonnet 5.5: output generation is more than 30 percent faster than Sonnet 5, and the model needs fewer tokens to complete the same work, cutting per-task cost by up to 30 percent even though the list price did not move. Remember that every vendor benchmarks its own models, so treat these as starting points, not verdicts. But the pattern is consistent: the mid-tier models are designed to win on completed work per rupee, not on raw per-token price.
Cache economics are the real battlefield
For agent workloads that replay the same documents and tool definitions across many runs, cache pricing can matter more than the headline rate. The numbers vary sharply between providers:
| Provider | Cache read (per million) | Cache write (per million) |
|---|---|---|
| OpenAI (GPT-6.1 Sol) | $0.10 | $2.50 |
| Anthropic (Sonnet 5.5) | $0.20 | $2.50 |
| Anthropic (Opus 5.5) | $0.20 | $5 |
| Google (Argon, introductory) | 95% off input (~$0.10) | not disclosed |
OpenAI’s $0.10 cached input rate is 95 percent below its standard input price, and half of GPT-6 Sol’s cached rate. For Indian developers building agent workloads with repeated context, that cache line alone is a meaningful cut: cached input at roughly Rs 9.60 per million tokens.
Read the footnotes before you budget
Three footnotes complicate any simple table. First, context bands: OpenAI charges $4 input and $15 output for GPT-6.1 Sol prompts above 272,000 tokens, so long-context work costs more than the headline rate. Second, effort settings change the bill. Anthropic notes that Claude products default to Medium effort while the Claude API defaults to High, so benchmark and cost comparisons made at different effort levels distort easily. Third, availability: Argon is not broadly on sale yet. It is rolling out first to trusted cyber defenders through Google’s Fairwind Program, and Google’s own developer pricing pages still list no Gemini 4 model. The $2 price is a promise about a product you cannot buy yet.
What this means: which tier for which job
A practical reading of the list, from cheapest to dearest:
GPT-6 Luna ($0.10/$0.50) is for high-volume, low-stakes work: classification, extraction, summarization at scale. When a task is simple and you are doing it a million times, Luna is the only tier where the arithmetic stays sane.
GPT-6.1 Sol, Claude Sonnet 5.5, and Gemini 4 Argon at $2/$10 are the new default for coding agents and professional workflows. Sonnet 5.5 and GPT-6.1 Sol score near flagship levels on several agentic benchmarks at one-fifth the per-token price of Astra. If you are running a coding agent today, this tier is where the testing should start.
Claude Opus 5.5 at $4/$20 remains the escalation tier for ambiguous work that needs sustained judgment. Anthropic says complex, open-ended tasks are still better suited to Opus. The 40 percent effective saving over Opus 5’s list rates softens the step up, but it is a step up.
GPT-6 Astra at roughly $10/$50 is for when accuracy is worth the multiple: the hardest science and engineering problems, where a failed task costs more than the tokens. Just note that the model it would have been succeeded by, GPT-6.1 Astra, was shelved in September after failing OpenAI’s own safety bar, so Astra stays the flagship until that changes.
The price war’s real winner so far is the mid-tier. A year ago, near-frontier capability at $2 per million input tokens was unthinkable. Now three labs ship it, and the only thing more volatile than the prices is how long the discounts last.
Related reading: Google Unveils Gemini 4 Argon: Benchmark Lead, a $2 Price Tag, and a Catch | GPT-6.1 Sol Arrives at DevDay: OpenAI’s $2-a-Million-Token Upgrade That Nearly Matches Astra | Claude Opus 5.5 Arrives: Anthropic’s New Flagship Costs 40% Less and Codes Faster
