GPT-6 Sol vs Claude Opus 5.5 — Which to Route To After the September Price War

Prices and benchmark figures checked on 26 September 2026 against the sources at the end. Benchmark numbers are vendor-reported or third-party runs, as labelled. None of them replaces a replay on your own tasks.

On 22 September 2026 Anthropic released Claude Opus 5.5. About 90 minutes later OpenAI released GPT-6 Sol and GPT-6 Luna. Fortune called it a price war: both labs cut prices on the models businesses actually run day to day. If you route API traffic between the two providers, the default you set this month decides most of next quarter's bill.

The short answer: GPT-6 Sol is half the price per token. Opus 5.5 leads on most public benchmarks, by a margin that varies a lot by task. Route by task class, and measure cost per successful task rather than cost per token.

Price per token

Two things stand out. First, Sol is exactly half Opus 5.5 on fresh input and on output. Second, cached reads cost the same on both, $0.20 per million. Agent loops that re-read a large, stable context narrow the gap, because the cached share costs the same on either model.

Both prices are cuts. GPT-6 Sol is 50% below GPT-5.6 Sol's promotional $4/$20. Opus 5.5 is 20% below Opus 5 per token and 60% below it on cache reads. Anthropic says Opus 5.5 also uses fewer tokens per task than Opus 5, so typical workloads cost about 40% less.

A worked cost-per-task example

Take one agentic coding task that reads 200,000 input tokens, 80% of them cache hits, and writes 20,000 output tokens. Assume both models use the same number of tokens (an assumption, not a measurement):

Opus 5.5 still costs about 1.9× per attempt, not 2×, because the cached share is priced the same. Cache writes ($2.50 per million on Sol, $5 on Opus) add a little on the first pass.

Now add the part the price list leaves out. Suppose, purely for illustration, that Sol passes your validator 70% of the time and Opus 5.5 85%, and that each failure costs an engineer 10 minutes (about $10 of time):

Change the inputs and the answer flips. At equal pass rates, Sol wins on cost every time. The lesson is that human time on failures dominates both token bills. Measure pass rates on your own tasks before you pick a default.

What the public numbers say

These are the comparisons available as of 26 September. Each uses a different setup, so compare within a row, not across rows.

• Artificial Analysis Intelligence Index (reported by MindStudio, 24 Sep): Opus 5.5 scores 51 at medium effort and 58 at max. GPT-6 Sol scores 40 at medium and 48 at max. The reported cost per evaluation task at medium effort was about $1.34 for Opus 5.5 and $0.25 for Sol, roughly 5.4× apart. So Opus 5.5 leads, at several times the cost per task.
• CursorBench 4.0 (Cursor's own table): Opus 5.5 at medium effort scores 52.5% at $2.91 per task, and at max effort 57.8% at $13.43. GPT-6 Sol had no row on the table when we checked. The best GPT-5.6 Sol row was 41.7% at $8.23.
• AutomationBench (business workflows across apps): OpenAI reports GPT-6 Sol at xhigh effort scoring 33.2% at $0.27 per task, ahead of Opus 5 at max (26.9%). Anthropic reports Opus 5.5 at 40.0% from Zapier's early-access run, with safeguard fallbacks counted as failures. These are different runs, so read them as "both improved on their predecessors" rather than a head-to-head.
• Hands-on app builds (MindStudio): one creator's eight-task test at medium effort scored Opus 5.5 75/80 and Sol 66/80. They tied on four tasks. Opus pulled ahead on interaction-heavy and multi-feature builds.

Where each one wins

Route to GPT-6 Sol when:

• The task is well specified and your validator catches failures cheaply (structured extraction, tool calls with schemas, standard CRUD code).
• Volume is high enough that a 2× token price shows up in the monthly bill.
• You are already on GPT-5.6 Sol. Moving to GPT-6 Sol halves the price with no provider change. See Migrating from GPT-5.6 to GPT-6 Sol and Luna.
• Your team works in ChatGPT Work or Codex, where GPT-6 Sol is available on paid plans.

Route to Claude Opus 5.5 when:

• The work is long-horizon and agentic (migrations, audits, multi-hour terminal work). This is where the published gaps are widest.
• Failures are expensive to catch: security-sensitive code, customer-facing documents, financial models.
• You need zero data retention. Anthropic says Opus 5.5 is available with ZDR, like previous Opus models. Unlike Fable 5.1, it is not a Covered Model with 30-day retention.
• Your team runs Claude Code, where Opus 5.5 also came with higher five-hour subscription limits.

Watch-outs on each side. Opus 5.5 ships with safeguards similar to Fable 5.1's. Anthropic says most cybersecurity tasks are re-routed to Opus 4.8. Security teams should test their real workload rather than assume Opus 5.5 handles it end to end. It also no longer runs with thinking switched off. On the OpenAI side, GPT-6 Sol has a 1.05M-token context and 128K max output, but any request above 272K input tokens bills at $4/$15 for the whole request. Size long-document work with that step in mind.

A routing rule you can test this week

1. Default standard traffic to GPT-6 Sol. It is the cheaper frontier-class model, and most routine tasks pass on either.
2. Escalate to Opus 5.5 on failure or by task label. Label long-horizon agent tasks and high-stakes outputs as "hard" up front, and send validator failures up a tier.
3. Send volume work down, not sideways. Classification and summarisation belong on a budget tier such as GPT-6 Luna ($0.10/$0.50), not on either frontier model. Claude Opus vs budget flash covers that split.
4. Log cost per successful task per route for two weeks, then move the escalation threshold based on data.

If you want the choice scored against your own constraints (data retention, provider lock-in, budget ceiling), the AI Model Picker walks through them. For the harness side of the same decision, use the AI Coding Assistant Finder. The late-September pricing snapshot lists every other frontier row.

Bottom line

GPT-6 Sol halves the per-token price of frontier-class work. Claude Opus 5.5 is the stronger model on most published measures, especially long agentic tasks, at twice the token price and a larger multiple per benchmark task. Neither headline answers the routing question on its own. Default to Sol, escalate to Opus 5.5 where your own pass rates justify it, and let two weeks of cost-per-success data move the line.

Sources (accessed 2026-09-26)

• OpenAI — Introducing GPT-6 Sol and Luna ($2/$10; 90% cached-input discount; AutomationBench 33.2% at $0.27/task; availability): https://openai.com/index/introducing-gpt-6-sol-and-luna/
• OpenAI — model docs for gpt-6-sol (1.05M context, 128K max output): https://developers.openai.com/api/docs/models/gpt-6-sol
• OpenAI — API pricing (gpt-6-sol cached $0.20, cache writes $2.50, >272K $4/$15, Fast mode $4/$20): https://developers.openai.com/api/docs/pricing
• Anthropic — Introducing Claude Opus 5.5 ($4/$20; cache reads $0.20, writes $5; fast mode; CursorBench 4.0 and AutomationBench results; ZDR; safeguards): https://www.anthropic.com/claude-opus-5-5
• Cursor — CursorBench 4.0 public table (Opus 5.5 rows and cost per task): https://cursor.com/cursorbench
• MindStudio — "GPT-6 Sol vs Claude Opus 5.5: Pricing Per Million Tokens Compared" (24 Sep 2026; Artificial Analysis index figures; hands-on test): https://www.mindstudio.ai/blog/gpt-6-soul-opus-5-5-pricing
• The Batch (DeepLearning.AI) — Claude Opus 5.5 (1M input, 128K output; Artificial Analysis and Vals rankings): https://www.deeplearning.ai/the-batch/claude-opus-5-5-leaps-forward
• Fortune — "OpenAI, Anthropic release dueling models as price wars heat up" (22 Sep 2026): https://fortune.com/2026/09/22/what-ai-slowdown-openai-anthropic-release-dueling-moreaffordable-models-as-ai-price-wars-heat-up/
• TechCrunch — OpenAI launches GPT-6 Sol and Luna (Opus 5.5 released 90 minutes earlier): https://techcrunch.com/2026/09/22/openai-launches-gpt-6-sol-and-luna/
• Claude Help Center — Covered Models (Fable and Mythos; Opus not listed): https://support.claude.com/en/articles/15425695-covered-models