Claude Opus 5.5 vs GPT-6.1 Sol vs Sonnet 5.5: we re-priced 3 automations after the September cuts.

For most judgement steps, GPT-6.1 Sol. It has the same sticker price as Claude Sonnet 5.5 but costs less on every shape we priced, and about 38% less on agent runs that re-read a long context. Use Claude Opus 5.5 only for the one step where being wrong is expensive, and on that step it is now half the price of GPT-6 Astra.

How we got these numbers. Every price on this page comes from Anthropic's API pricing page and OpenAI's API pricing page, both read on 7 October 2026. The workflow shapes, token counts and arithmetic are ours, and every assumption is written out so you can swap in your own. We have not run a quality benchmark between these models, and we will not pretend to. This is a cost comparison with an opinion attached, not a leaderboard.

Four calls drive the verdict:

• Same sticker, different bill. GPT-6.1 Sol and Claude Sonnet 5.5 both list at $2 in and $10 out. Sol still comes out cheaper because its cached input is half the price, OpenAI charges no extra to write the cache, and Anthropic says its newer models produce about 30% more tokens for the same text.
• Opus 5.5 is a straight saving over Opus 5. Same family, 20% lower list price, 60% cheaper cache reads. If any workflow calls Opus 5 today, switching the model name is the cheapest win on this page.
• GPT-6 Astra lost its job. On our agent run, Opus 5.5 costs $437 per 1,000 runs and Astra $870. GPT-6.1 Sol is $153. Astra now needs to be meaningfully better than both on your task to earn a place.
• The cheap tier still does most of the work. A triage step costs $0.30 per 1,000 runs on GPT-6 Luna and $6 on GPT-6.1 Sol. Nothing in September changes the rule: start on the cheapest model and escalate only what fails.

Per 1,000 runs, a triage step costs $6 on GPT-6.1 Sol, $7.80 on Sonnet 5.5 and $15.60 on Opus 5.5. A document extraction costs $34, $44.20 and $88.40. A cache-heavy agent costs $153, $245.70 and $436.80.

These are the three shapes behind most of the n8n, Make and Zapier AI steps we build for small teams:

• A. Triage and classify. Read a ticket, lead or email and label it. 2,000 input tokens and 200 output tokens per run, no caching.
• B. Document extraction. Pull fields out of an invoice, contract or form. 12,000 input and 1,000 output per run, no caching, because every document is different.
• C. Multi-step agent. An agent loop of 8 turns over a 30,000-token context (instructions, tool definitions, a knowledge doc). The context is written to the cache once and read back on the other 7 turns, plus 2,000 fresh input and 500 output tokens per turn.

Assumptions you can change: list prices, standard tier, no batch discount, and Claude Opus 5.5 and Sonnet 5.5 token counts raised 30% for Anthropic's tokenizer note. Claude Haiku 4.5 is an older model on the previous tokenizer, so it is not adjusted.

Read the columns, not the rows. On triage, the whole spread from GPT-6.1 Sol to Opus 5.5 is under $10 per 1,000 runs, so pick on quality, because the money barely moves. On the agent shape the same choice is worth $284 per 1,000 runs. That is where the September price cuts actually pay out, and it is almost entirely a caching story.

Even if you ignore the tokenizer adjustment completely, the ranking holds. At face value Sonnet 5.5 comes to $189 on the agent shape, still 24% more than GPT-6.1 Sol's $153.

Over Opus 5, yes, with no catch: it is 20% cheaper on simple steps and 29% cheaper on our cached agent at the same token counts. Over GPT-6 Astra, on price alone, yes: it is about half the cost on every shape we priced.

Anthropic's launch claim is that typical workloads cost about 40% less on Opus 5.5 than on Opus 5, as Inc42 reported at launch. Our arithmetic only reaches 20% to 29%, and the gap is worth understanding. The lower list price and the cheaper cache reads account for everything we can calculate. The rest of Anthropic's 40% rests on Opus 5.5 using fewer tokens to finish the same task, which is a vendor claim about behaviour, not a price. You will only see it in your own run logs.

So our advice is simple. Any step on Opus 5 should move to Opus 5.5 now, because even with zero token savings it is cheaper. Any step on GPT-6 Astra should be tested against Opus 5.5 and GPT-6.1 Sol on 20 of your real examples. Keep Astra only if it clearly wins on the cases that fail elsewhere. OpenAI pitched GPT-6.1 Sol at DevDay on 29 September as near-Astra intelligence at a fifth of Astra's token price, so the burden of proof now sits with Astra.