September 25, 2026. The flagship fight gets the attention, but agencies buy in the middle of the range. That is where OpenAI's GPT-6 Sol, released September 22 at $2 per million input tokens and $10 output, now meets Anthropic's Claude Opus 5.5, released the same day at $4 and $20. OpenAI positions Sol as its model for complex coding and agentic workflows; Anthropic positions Opus 5.5 as its model for long-running agentic coding and knowledge work. For the intake agents, CRM automations and document pipelines we build for SMBs, these two are the realistic shortlist, and the choice turns on three things: the per-token price, what happens to that price when a request gets long, and which vendor's business-workflow numbers you believe.
The numbers that decide it
- List price. Sol is exactly half of Opus 5.5 on both input ($2 against $4) and output ($10 against $20).
- Cache. Both charge $0.20 per million for cached input. Cache writes are $2.50 on Sol and $5.00 on Opus 5.5 for the short cache; Anthropic also sells a one-hour cache at $8.00, and OpenAI's caching guide offers extended retention of up to 24 hours as a request option, with no separate write price listed on its pricing table.
- Long context. OpenAI's Sol model page states that prompts over 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request, so Sol becomes $4, $0.40 and $15 past that line. Anthropic's pricing page bills Opus 5.5's full 1M window at standard rates. At 400,000 input tokens and 4,000 output, Sol costs $1.66 and Opus 5.5 costs $1.68, so the price gap closes to nothing.
- Fast tiers. Sol's Fast mode is $4 and $20; Opus 5.5's is $8 and $40. Batch halves both: Sol to $1 and $5, Opus 5.5 to $2 and $10.
- Context and cutoff. Both are 1M tokens with 128K output. Sol's knowledge cutoff is April 20, 2026; Opus 5.5's is June 2026.
The business-workflow benchmark both sides quote
Zapier's AutomationBench tests agents on end-to-end business workflows across 47 tools in sales, marketing, operations and support, which makes it the one public test closest to what an agency actually ships. OpenAI's launch page says GPT-6 Sol at xhigh effort outperforms Claude Opus 5 at max effort at 9% of Opus 5's cost per task. Anthropic's launch page prints Zapier's leaderboard: Opus 5.5 at 40.0%, GPT-6 Astra 41.4%, GPT-5.6 Sol 28.8%, Opus 5 26.9%. Neither page prints GPT-6 Sol against Opus 5.5 on the same run, and Anthropic notes its own score was taken with safeguard interventions counted as failures. So the honest reading is: OpenAI's claim is against the previous Anthropic model, Anthropic's table is against the previous OpenAI mid-tier, and the two new models have not yet been compared head to head by a third party.
What it means for operators
For short, high-volume turns, which is what an AI receptionist, a lead qualifier or a form-filling agent produces all day, Sol's half price is real and compounds: a 50,000 token turn with 40,000 cached and 5,000 output costs about $0.08 on Sol and $0.15 on Opus 5.5. For long documents and long-running agents, the gap disappears past 272K tokens on Sol, and Opus 5.5's flat pricing plus Anthropic's efficiency claims (fewer steps, fewer output tokens, faster generation than Opus 5) make it the safer default. The sensible pattern for a client stack is to route by shape rather than pick a winner: Sol for the conversational front end, Opus 5.5 for the document and audit work, and a usage log that records tokens per task so the split can be checked monthly. That is how we size a build like an AI calling agent for insurance agencies, where the call itself is short and cheap but the policy documents behind a quote are not, and it is the kind of routing decision an AI automation partner should make in the design, not after the first invoice.
Frequently Asked Questions
On list price, yes: $2 input and $10 output per million tokens against $4 and $20. Cached input is $0.20 on both. Past 272K input tokens Sol's whole request bills at $4 input and $15 output, at which point a 400K token request costs about the same on both models.
There is no third-party head-to-head yet. OpenAI says GPT-6 Sol at xhigh effort beats Claude Opus 5 on Zapier's AutomationBench at 9% of the cost per task; Anthropic prints Zapier's leaderboard with Opus 5.5 at 40.0% against GPT-5.6 Sol at 28.8%. Each vendor compares against the other's previous model.
Both offer a 1M token context window and 128K output tokens. Sol's context is billed in two tiers with the higher rate applying to the whole request over 272K input tokens; Opus 5.5 is billed at standard rates across the full window.
Route by workload shape: Sol for short, high-volume conversational turns where its half price compounds, Opus 5.5 for long documents and long-running agents where flat context pricing and token efficiency matter. Log tokens per task and review the split monthly.