Claude Opus 5.5 vs GPT-6 Sol — Cost, Performance, Coding Agents

On September 22 Anthropic shipped Opus 5.5, and about an hour and a half later OpenAI shipped GPT-6 Sol and Luna. Not a coincidence.

Like the Astra post, I waited two days. On launch day each vendor only had its own table, and each table compared against the other side’s previous model. Anthropic’s has GPT-5.6 Sol. OpenAI’s has Opus 5. Neither included the new one. Since then Artificial Analysis has published measurements, so the two can finally sit side by side.

This post looks at three things only: cost, overall performance, and coding agents.

Claude Opus 5.5

claude-opus-5-5

Input

$4 / MTok

Output

$20 / MTok

Cache read

$0.20 / MTok

Context

1M

GPT-6 Sol

gpt-6-sol

Input

$2 / MTok

Output

$10 / MTok

Cache read

$0.20 / MTok

Context

1.05M

Specs

Sol is exactly half in almost every row. Cache reads are the exception.

Opus 5.5 cut per-token prices by 20% and cache reads by 60% versus Opus 5. Sol simply halved GPT-5.6 Sol’s price. Both got cheaper; OpenAI cut deeper.

Luna, released alongside Sol, is $0.10 / $0.50. Different weight class, so it’s left out here.

Cost: half price is not half price

On the price sheet Sol is always half. In practice it isn’t, for two reasons.

Cache reads cost the same

Agents re-read the same context over and over, so most of their tokens are cache reads. That row is $0.20 on both. The more cache-heavy the workload, the smaller the gap.

Sol has a surcharge above 272K

Once input goes over 272K tokens, the whole request is billed at 2x input and 1.5x output. Opus 5.5 has no surcharge up to 1M.

Digital Applied priced the same token mix on both: 400K input, 8M cache reads, 600K cache writes, 300K output.

Inside the surcharge band Sol ends up slightly more expensive. Even without it, Sol is about 57% of Opus, not 50%. If your agent loads a large repo into context, run this math first.

One more thing. On Opus 5.5, changing top-level effort mid-conversation invalidates the cache. Sol reportedly keeps the cache across effort changes and tool toggles. Anthropic has a per-message effort option, but it’s still in beta.

Cost per task by effort

This matters more than the per-token price. Artificial Analysis Intelligence Index (v4.3.2), run at every effort level:

At the same effort, Sol is roughly 5x cheaper per task. The token price gap is 2x, so Opus 5.5 is clearly spending a lot more tokens.

Line them up by dollars instead and it changes:

• Around $0.50: Sol xhigh 44.1 ($0.53) vs Opus low 42.3 ($0.55). Same money, Sol slightly ahead
• Around $1: Sol max 47.5 ($1.06) vs Opus medium 51.2 ($1.34). 30 cents more buys 4 points on Opus

Roughly: below 50 cents per task Sol wins, above a dollar Opus wins. Sol tops out at 47.5 even at max, which is below Opus’s default.

Opus has its own diminishing returns. Going from xhigh to max adds 1.6 points for $2.50 more. Max looks like a benchmark setting more than a daily one.

Overall performance

Vendor tables

As mentioned, neither vendor compared against the other’s new model. Pulling out the rows that overlap:

Opus 5.5 leads both. But as noted in the Astra post, OSWorld numbers from different task releases aren’t directly comparable. I wouldn’t take the 21-point gap at face value.

The rest of Anthropic’s table: Terminal-Bench 4.0 66.4%, FrontierCode 54.4%, CursorBench 57.8%, HLE 67.7%, GDPval-AA 1846. All at max effort.

OpenAI’s table: DeepSWE 68.8%, Agents’ Last Exam 56.4%. The headline line was “within 1.1 points of Fable 5 xhigh (69.9%) at 80% lower cost.”

Some reviewers also pointed out that on OpenAI’s own DeepSWE and OSWorld charts, GPT-5.6 Sol scores higher than GPT-6 Sol. Halving the price seems to have cost something on some workloads.

Third party: Artificial Analysis

A 10-point gap. Not small.

Opus 5.5 is now #1 on the AA index, above Fable 5.1. Anthropic’s “Fable 5.1-class work, cheaper” pitch holds up here.

Per evaluation (Opus at medium vs Sol at max):

Opus at medium still beats Sol at max on most rows.

The last two rows are the interesting ones. Opus gets more answers right, but when it doesn’t know, it makes things up more often. Sol dropped from 92% to 60% hallucination versus GPT-5.6 Sol. It looks tuned to say “I don’t know” more.

Coding agents

The part I actually cared about.

AA Coding Agent Index (v1.5)

Each model in its own harness at max effort.

9 points apart, at 4.4x the price.

Breakdown:

The terminal gap is big: 20 points. DeepSWE looks even, but Sol’s number is OpenAI’s own run in a different harness, so I wouldn’t call it a tie.

Also: in the Astra post I listed Fable 5.1 at 70 on this index. Under v1.5 it’s 62. Don’t compare scores across index versions. I say this every time and still trip on it.

Opus 5.5 got cheaper per token and more expensive per task

Opus 5 to 5.5 cut the token price by 20%, yet cost per task on the coding index went from $10.79 to $13.04, up 21%.

Output tokens per task went from 137K to 333K, 2.4x. It thinks longer and does more. That’s what bought the 6 extra points. It’s not cheaper; it does more work for similar money.

Anthropic’s “40% cheaper on typical workloads” is probably about the medium default. Run it at max in Claude Code and the bill may go up.

Sol gained 2 points over 5.6

Sol is +2 on the coding index versus GPT-5.6 Sol at half the cost per task. Terminal-Bench 4.0 went 37% → 43%, SWE-Atlas-QnA 54% → 58%.

This isn’t a release that raises the ceiling. It sells the same performance for half the price, which is also how AA summarized it.

Vendor case studies

All vendor-selected, so read accordingly.

Opus 5.5:

• A 680,000-line code migration finished in under a day
• Audit and repair of a 200,000-line codebase in under three hours; Opus 5 took 20+ hours and 2.5x the tokens
• HAProxy C → Rust in 9.5 hours versus 12 for Fable 5.1, at 51% lower cost
• On FrontierCode, medium effort beat Astra at max for about 20% of the cost

Sol:

• GitHub Copilot saw a 50%+ drop in freshly processed prompt tokens thanks to caching changes
• OpenAI says it states what it verified and what it didn’t, skips the preamble, and leads with the diff

One hands-on test

DataCamp gave both the same prompt: a single-file HTML Tetris clone where gravity flips every 20 seconds, at high effort.

Both got the gravity flip right. Sol won on completeness and UI polish. Sol iterates in small steps; Opus writes big chunks in one go. Opus used fewer tokens but cost more than 3x because of the price.

One test doesn’t generalize. It does support the idea that for small, self-contained tasks, Sol is enough.

Harnesses

Both models run best in their own agent: Opus 5.5 in Claude Code, Sol in Codex. Both are selectable in GitHub Copilot.

Opus 5.5 is now the default in Claude Code, the Claude app and Cowork for Pro / Max / Team. I run this blog through Claude Code and it had already switched over.

On the Codex side there’s an open issue: Codex CLI clamps Sol’s effective context to 828.4K, against a 1.05M spec. GPT-5.6 Sol had similar context-clamping issues before.

Plus plan limits are 15–150 Sol messages per five hours, up from 10–100 on 5.6 Sol. On HN a fair number of people said they liked 5.6 Sol’s communication style and engineering instincts, and some feel 6 doesn’t quite have it.

What breaks when you switch

Opus 5 → Opus 5.5

1. Thinking can’t be disabled. Both {type: "disabled"} and budget_tokens return 400. Effort is the only knob
2. Default effort is medium. Opus 5 defaulted to high. Swap only the model name and you run one level lower. Set it explicitly
3. tool_choice any / tool returns 400. Use auto with strict: true and steer from the prompt, or structured outputs
4. Thinking blocks are tied to the model and conversation. A fallback to Opus 5 runs without them
5. Computer use only via computer_toolset_20260801. computer_20251124 returns 400
6. Text between tool calls now arrives as progress-update thinking blocks. Set display: "updates" to see it

Number 2 is the quiet one. No error, just lower scores.

GPT-5.6 Sol → GPT-6 Sol

1. Tool calls in Chat Completions only work with reasoning_effort: none. For reasoning plus tools, use the Responses API
2. Effort ranges from none to max, default medium
3. Requests over 272K get the surcharge. Recalculate if you fill the context

Sol can turn reasoning off, which matters for classification and extraction. Opus 5.5’s floor is low, so even trivial calls carry some thinking cost.

Safety, briefly

Opus 5.5 shipped with a system card and an external METR evaluation. Anthropic says attempts to circumvent boundaries dropped about 85% versus Opus 5. The classifiers got broader, adding bio and reasoning_extraction categories, so false-positive refusals may tick up.

Sol and Luna launched without system cards. Sol attempted unauthorized workarounds in 64.4% of warning-compliance runs, barely down from 5.6 Sol’s 68.2%. Some users also report the new security classifier pausing long autonomous runs to ask for confirmation.

Summary

Opus 5.5 is clearly stronger overall. 10 points on the AA index, 9 on the coding index. Sol at max doesn’t reach Opus at medium.

Sol is clearly cheaper. About 5x per task at the same effort, 4.4x on the coding index. But with heavy caching or past 272K, most of that gap disappears.

For coding agents it depends on task size. Terminal-heavy work, large repos and long runs favor Opus 5.5. Small self-contained tasks and high-volume runs favor Sol. On the single-file Tetris test, Sol produced the better result.

Opus 5.5’s “40% cheaper” is a medium-effort number. At max it costs more per task than Opus 5.

My own default would be Opus 5.5 at medium for coding, with repetitive jobs split off to Sol. Max only when genuinely stuck.

Both models are two days old. Sonnet 5.5 and Haiku 5.5 are supposedly a few weeks out, so this will need another look then.

References

• Anthropic announcement — https://www.anthropic.com/claude-opus-5-5
• OpenAI announcement — https://openai.com/index/introducing-gpt-6-sol-and-luna/
• Artificial Analysis on Opus 5.5 — https://artificialanalysis.ai/articles/claude-opus-5-5
• Artificial Analysis on Sol / Luna — https://artificialanalysis.ai/articles/gpt-6-sol-and-luna-push-the-cost-efficiency-frontier
• Artificial Analysis comparison — https://artificialanalysis.ai/models/releases/comparisons/gpt-6-sol-vs-claude-opus-5-5
• Coding Agent Index breakdown — https://www.orcarouter.ai/blog/claude-opus-5-5-coding-agent-index
• Cost per task breakdown — https://www.digitalapplied.com/blog/gpt-6-sol-vs-claude-opus-5-5-cost-benchmarks
• Effort ladder comparison — https://codingfleet.com/blog/claude-opus-5-5-vs-gpt-6-sol/
• DataCamp Tetris test — https://www.datacamp.com/blog/gpt-6-sol-vs-claude-opus-5-5
• Vellum on Sol benchmarks — https://www.vellum.ai/blog/gpt-6-sol-and-luna-benchmarks-explained
• VentureBeat on Opus 5.5 — https://venturebeat.com/technology/anthropic-releases-claude-opus-5-5-beating-fable-5-1-on-key-agentic-benchmarks-at-60-cheaper-api-price
• SiliconANGLE — https://siliconangle.com/2026/09/22/anthropic-releases-claude-opus-5-5-and-openai-counters-with-two-cheaper-gpt-6-models/
• Codex context issue — https://github.com/openai/codex/issues/47805