The October 2026 Model Price Sheet: Astra, Fable 5.1, Opus 5.5, GPT-6.1 Sol, Sonnet 5.5 and Gemini 4 Argon, With a 10,000-Ticket Bill

The one-paragraph version

If you are building a support or back-office agent and have not chosen a model, the mid tier is a near-tie on list price: Claude Sonnet 5.5 and GPT-6.1 Sol are $2 per million input tokens and $10 per million output; Google has announced Gemini 4 Argon at the same introductory price, but it is not yet sold to API customers. Pick on quality for your task and on three second-order prices: what a cached input token costs, whether long prompts carry a surcharge, and whether the price you see is introductory. Then put two dates in the calendar: 30 November 2026 (Claude Sonnet 4.5 retires) and 1 January 2027 (Gemini Flash prices double).

The sheet

Prices per million tokens. Prices and plan names are as published on each vendor's own pricing page; last verified 3 October 2026. "Cache read" is the price of an input token served from the provider's prompt cache.

Three structural notes that matter more than any single row:

Cache reads are where the labs actually compete. A support agent sends the same system prompt and knowledge snippets on every turn. Anthropic prices a Fable 5.1 cache hit at 2.5% of input and an Opus 5.5 hit at 5%; OpenAI's GPT-6.1 Sol hit is 5% and Astra's is 10%; Google has announced 95% off for Argon. On a heavily cached workload these ratios move the bill more than the headline input price does.

Long context is a surcharge at OpenAI and not at Anthropic. OpenAI's page defines long context as more than 272K input tokens and roughly doubles the rate. Anthropic's page says Claude 4.6 and later include "the full 1M token context window at standard pricing." If your agent reads whole documents, that is a real difference.

Tokenizers differ. Anthropic states that Claude 4.7 and later use a tokenizer that "produces approximately 30% more tokens for the same text" than Claude Sonnet 4.6 and earlier. If you are migrating off Sonnet 4.5 and comparing against your current token counts, multiply the Anthropic rows by roughly 1.3 before you compare. Per-token prices across labs are not directly comparable without measuring your own text.

The worked bill: 10,000 support tickets a month

Assumptions, stated so you can change them: each ticket is one model call with 3,000 input tokens, of which 2,000 are a cached system prompt and knowledge block and 1,000 are fresh (the customer message and retrieved context), plus 700 output tokens. That is 10 million uncached input, 20 million cached input and 7 million output tokens a month. Cache-write charges are ignored (they recur once per cache window, not per ticket). No Batch discount, because support is interactive. Anthropic's own worked example on the same page lands at "~$37.00 per 10,000 tickets" on Haiku 4.5 with a different token mix; the point of the table is the ratios, not the cents.

Arithmetic for one row, so you can check the rest: Sonnet 5.5 is 10M × $2 + 20M × $0.20 + 7M × $10, divided by one million, which is $20 + $4 + $70 = $94.

What the table says:

• At the mid tier the two models you can actually buy are within $2 of each other on this workload, and Argon is announced at the same point. Model choice there is a quality decision, not a price decision.
• The Flash cliff is a 2x on the exact same traffic. A team on 3.8 Flash at $35.25 a month with caching ($48.75 without) wakes up on 1 January at $70.50 ($97.50 without), which is close to Sonnet 5.5 money without the model change.
• When Argon reaches API customers, its introductory price buys Sonnet-class economics and its standard price is Opus 5.5 economics. Google's post gives neither a developer availability date nor the end of the introductory period. Budget at $184, enjoy $92.
• The frontier tier is five times the mid tier. Astra and Fable 5.1 are for the tickets your mid-tier model gets wrong, routed to them deliberately, not for every ticket.

The two dates

30 November 2026. Anthropic announced on 30 September that Claude Sonnet 4.5 is deprecated, with retirement on the Claude API scheduled for 30 November 2026. Anything pinned to that model string returns errors after that date (Anthropic's August retirement of Opus 4.1 worked the same way). The migration is to Sonnet 5.5 at a lower list price, with the 30% tokenizer caveat above, and it needs an eval run, not a string replace. The agent evaluation and regression harness is the workflow for exactly that.

1 January 2027. Gemini 3.7 and 3.8 Flash go from $0.75 to $1.50 input and $3.75 to $7.50 output. That is printed on Google's pricing page. If Flash is your default cheap tier, decide before December whether the January price still beats Haiku 4.5 at $1 / $5 or GPT-6 Luna at $0.10 / $0.50 for your task.

Two things the sheet does not show

Platform fees. The sheet is the token price only. If you run the model inside a hosted agent runtime, add the runtime's own meter. Anthropic's Claude Managed Agents adds $0.08 per session-hour while a session is running; OpenAI's hosted containers run $0.03 to $1.92 per 20-minute session depending on container size. Our enterprise agent platform comparison works those through.

Regional inference. Anthropic charges 1.1x for US-only inference on Claude 4.6 and later. Bedrock and Google Cloud regional endpoints carry a 10% premium over global endpoints. If data residency is a requirement, it is also a line item.

Next step

Price your own workload, not mine: pull a week of real tickets, count tokens with each provider's counting endpoint, and rerun the table. Then build the routing so the expensive model only sees the tickets that need it; the RAG support agent workflow and support inbox triage are the two halves of that. For the surrounding decisions, read how AI agent pricing models work and the free AI Agents Starter Kit; enterprise teams should go to the Enterprise AI Agent Platform Buyer's Guide. Head-to-heads: Claude API vs. OpenAI API, Claude API vs. Gemini, Gemini vs. OpenAI API. Tool pages: OpenAI API, Claude API, Gemini. For the enterprise view of the same decision, start at the enterprise hub.