Claude Opus 5.5 vs GPT-6 Sol: Price, Benchmarks, Coding (2026)

Anthropic shipped Claude Opus 5.5 on 22 September 2026. About 90 minutes later OpenAI released GPT-6 Sol and its budget sibling GPT-6 Luna, according to TechCrunch. The two releases target the same buyer: a team running coding agents and long agentic workflows through an API, where both quality and the bill matter.

The two models are not the same class, though. Opus 5.5 is Anthropic's recommended default for most workloads. GPT-6 Sol is OpenAI's mid-tier model, sitting below GPT-6 Astra. This comparison covers pricing, cost per task, the benchmarks both vendors actually share, and how each fits into Claude Code and Codex. For everything else about OpenAI's release, see our GPT-6 Sol and Luna guide. For Anthropic's side, see the Claude Opus 5.5 complete guide.

How do Opus 5.5 and GPT-6 Sol compare at a glance?

The spec sheets are closer than the prices suggest. Both have a context window of about 1M tokens, 128K max output, text-and-image input, and a reasoning effort dial that defaults to medium. The main structural differences are that Sol can switch reasoning off (none) and Opus 5.5 cannot, and that Sol reprices prompts longer than 272K tokens.

Sources: Anthropic's Opus 5.5 model page and OpenAI's GPT-6 Sol model page. Artificial Analysis lists Sol's usable context as 872K tokens, which is lower than OpenAI's figure. In practice, plan around OpenAI's documented 922K input limit.

How much do Claude Opus 5.5 and GPT-6 Sol cost?

The widely reported $2/$10 for Sol is correct, and OpenAI says it is permanent pricing, not a launch promotion. Opus 5.5's $4/$20 is also confirmed, a 20% cut from Opus 5. The headline is that Sol costs half as much per token. The pricing pages add two details that change that picture.

Sources: Anthropic pricing, OpenAI pricing.

Cache reads cost the same. Anthropic gave Opus 5.5 an unusual 0.05x cache-read multiplier, so a cache hit costs $0.20 per million tokens. That is exactly Sol's cached-input price. Agent loops re-send the same system prompt, tool definitions and repo context on every turn, so most of their input is cache hits. For that kind of workload, Sol's input advantage is much smaller than "50% cheaper" suggests. What remains is fresh input, cache writes, and above all output.

Long prompts close the gap further. Once a Sol prompt goes past 272K tokens, input doubles to $4 and output rises to $15. Opus 5.5 bills the full 1M window at the standard rate. If you routinely load whole repositories or large document sets into one call, you will be paying Opus's input price on Sol anyway, and saving only 25% on output. For a primer on why caching dominates agent costs, see how prompt caching reduces AI cost.

Which is cheaper per task?

Sol is, by a wide margin, and the reason is output tokens as well as price. Artificial Analysis ran its Intelligence Index on both models at max effort:

Sources: Artificial Analysis model pages for Opus 5.5 and GPT-6 Sol, and OfficeChai's write-up for the Luna figures.

At max effort, Opus 5.5 produced more than three times as many output tokens as Sol, at twice the output price. That works out to roughly 5.6x the cost per task. It also bought 10 more index points, which is a large gap at the top of this leaderboard. Sol's 48 is one point above GPT-5.6 Sol's 47, so OpenAI's gain this generation is mainly price, not capability.

Max effort is the worst case for Opus 5.5. Anthropic's own CursorBench 4.0 figures put Opus 5.5 at 52.5% for about $3 per task at medium effort, and 56.0% for about $4 at high effort. At max effort it scores 57.8%, and Anthropic does not list a cost for that run. The Opus 5.5 system card also notes that its FrontierCode scores decline above medium effort. For Opus 5.5, going from medium to max effort costs a lot of money for a small gain in score.

A worked example: one agent session

Take a coding-agent session that sends 2M input tokens across its turns. 90% of them are cache hits, the other 200K are written to cache, and the session produces 60K output tokens. Every prompt stays under 272K.

With the same token counts, Sol costs about 43% less here, not 50%, because the cache-read line costs the same on both. Real sessions will not use the same token counts. The AA data suggests Opus 5.5 tends to use more output tokens, while Anthropic's customer reports claim it finishes tasks in fewer turns than Opus 5. Run your own workload through both before you commit to either price.

How do Opus 5.5 and GPT-6 Sol compare on benchmarks?

This is where most head-to-heads go wrong. Anthropic's Opus 5.5 benchmark table compares against GPT-6 Astra and the previous GPT-5.6 Sol, because GPT-6 Sol did not exist when it was written. OpenAI's Sol figures, as reported by VentureBeat, use a different set of benchmarks. The table below lists only benchmarks where a number exists for both models. It says who ran each one, because vendor harnesses differ.

Where the comparison is clean, Opus 5.5 wins. It leads the independent AA index by 10 points, and its self-reported DeepSWE and AutomationBench scores beat Sol's self-reported scores. OSWorld 2.0 is ambiguous. Sol's 60.5% is well above Opus 5.5's strict score and well below its partial-credit score, and without knowing OpenAI's scoring mode, no fair winner can be called. For context, OpenAI says Sol gets similar computer-use results to Claude Opus 5 at a much lower cost (9to5Mac).

Not published for GPT-6 Sol as of launch day: Terminal-Bench 4.0, SWE-bench Pro, FrontierCode, CursorBench, GDPval-AA and HLE. On those, Opus 5.5 scores 66.4%, 89.9%, 54.4%, 57.8%, 1846 Elo and 67.7% (with tools) respectively (Anthropic, system card). The only OpenAI mid-tier figures on those benchmarks are for the older GPT-5.6 Sol: 37.3% on Terminal-Bench 4.0, 41.7% on CursorBench and 47.5% on FrontierCode. Do not treat those as GPT-6 Sol numbers.

Not published for Opus 5.5: Agents' Last Exam, where Sol scores 56.4% at max effort per OpenAI. OpenAI's launch claim that Sol and Luna "handle tasks substantially better than Anthropic's top models — like Fable and Opus" (TechCrunch) refers to earlier Claude models. It predates Opus 5.5 and none of the published numbers back it up against Opus 5.5.

One more independent signal: Artificial Analysis rates Sol at 57 on its Coding Agent Index, at about $2.99 per task, up from 55 for GPT-5.6 Sol. We could not find a published Opus 5.5 score on that index at the time of writing. For how Opus 5.5 compares with OpenAI's flagship, see Opus 5.5 vs GPT-6 Astra vs Grok 4.7.

Which is better for coding agents: Claude Code or Codex?

For most developers the choice between these models comes with a choice of agent. Opus 5.5 runs in Claude Code. GPT-6 Sol runs in Codex. Each vendor tunes its model for its own agent. AA's coding-agent evaluations also show scores shifting with the harness, so a model's score in one agent is not a reliable guide to how it performs in the other.

Opus 5.5 in Claude Code is included on Claude's paid plans, and Anthropic raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise at launch. Effort runs from low to max, and there is a fast mode on the first-party API. The published strengths are long, multi-step engineering work. Anthropic cites a 200k-line codebase audit in under 3 hours, a 680k-line migration in under a day, and Stripe having one session direct a dozen others through a rebase of 40 stacked PRs. The weak spots are token use at high effort and stricter cyber safeguards that can reroute security work to Opus 4.8. We have not confirmed whether Claude Code now defaults to Opus 5.5. Check /model after updating.

GPT-6 Sol in Codex is rolling out to Plus, Pro, Business, Enterprise and Edu accounts. Enterprise admins must enable it first (9to5Mac). You can select it directly:

codex exec -m gpt-6-sol "Review the current changes and fix failing tests"

Sol's pitch is sustained agent use at low cost. OpenAI reports it discloses tool failures far more honestly than GPT-5.6 Sol, with an undisclosed failure rate of 5.4% against 77.8% (VentureBeat). That matters when an agent runs unattended. Early Hacker News reactions praised the price and instruction-following. Some commenters said Sol tends to over-engineer, and some said they miss 5.6 Sol's style.

Our rule of thumb: if your agents do long refactors, migrations or multi-repo work where a failed run costs more than the tokens, use Opus 5.5 in Claude Code. If you run many parallel, well-scoped tasks such as test fixes, lint cleanups, small features or PR reviews, Sol in Codex gets more done for the same budget. Our Claude Code vs OpenAI Codex comparison and the AI coding agents guide cover the agents themselves in more depth.

What are the API differences?

Moving code between the two takes more than swapping the model string. The differences that break integrations:

• Thinking can't be turned off on Opus 5.5. thinking: {"type":"disabled"} and fixed budget_tokens both return a 400 error. Control depth with output_config.effort. Sol supports reasoning.effort: "none" for fast, reasoning-free calls.
• Forced tool use is gone on Opus 5.5. tool_choice of type any or tool returns 400. Use auto with strict tool use or structured outputs instead.
• Text between tool calls now arrives as thinking blocks on Opus 5.5. At the default display setting those blocks are empty, so progress UIs that stream that text go quiet. Our Opus 5.5 migration guide walks through the fixes.
• Hosted tools are OpenAI's advantage. Through the Responses API, Sol can use web search, file search, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search without you running them.
• Distribution is Anthropic's advantage. Opus 5.5 ships on Bedrock, Google Cloud and Microsoft Foundry on day one, which matters if your procurement or data residency runs through a hyperscaler. Some trackers list Sol on Bedrock, but we could not confirm that in OpenAI's or AWS's documentation.

Minimal calls for each:

import anthropic

client = anthropic.Anthropic()
resp = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=16000,  # leave headroom: thinking counts against it
    output_config={"effort": "medium"},
    messages=[{"role": "user", "content": "Refactor this module for testability: ..."}],
)
# read blocks by type, not position
print("".join(b.text for b in resp.content if b.type == "text"))

from openai import OpenAI

client = OpenAI()
resp = client.responses.create(
    model="gpt-6-sol",
    reasoning={"effort": "medium"},  # "none" is also valid on Sol
    input="Refactor this module for testability: ...",
)
print(resp.output_text)

Where does GPT-6 Luna fit?

GPT-6 Luna (gpt-6-luna) is the budget option at $0.10 input and $0.50 output, with $0.01 cached reads. It has the same ~1M context as Sol and a 18 May 2026 knowledge cutoff. OpenAI positions it for high-volume, clearly scoped tasks such as summarising, extraction and quick answers. It reports 66.6% on DeepSWE 1.1, only 2.2 points behind Sol's 68.8%. Artificial Analysis scores it 37 on the Intelligence Index at $0.07 per task, level with GPT-5.6 Luna's 37. Free and Go ChatGPT users get Luna in the desktop app.

Luna is not an Opus 5.5 competitor. It is the model you put underneath one. A common pattern is to route triage, extraction and simple edits to Luna and escalate hard cases to Opus 5.5 or Sol. Anthropic's cheapest current options are Sonnet 5 at $2/$10, the same price as Sol, and Haiku 4.5 at $1/$5, which is still 10x Luna's price. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow "in the coming weeks".

Which should you pick?

If your team uses both, the practical setup is to route by task difficulty rather than pick one model. Send well-scoped work to Sol, send expensive-to-fail work to Opus 5.5 at medium or high effort, and track cost per merged change, not cost per token. Our earlier Opus 5 vs GPT-5.6 comparison shows how this pattern played out in the previous generation.

FAQ

Is GPT-6 Sol better than Claude Opus 5.5?

No, not on published evidence. Opus 5.5 scores 58 to Sol's 48 on the independent Artificial Analysis Intelligence Index. It also leads on DeepSWE and AutomationBench, the only coding and agent benchmarks where both vendors report a number. Sol's advantage is cost: about half the per-token price and roughly a sixth of the measured cost per task.

How much does GPT-6 Sol cost compared with Opus 5.5?

GPT-6 Sol costs $2 input and $10 output per million tokens. Claude Opus 5.5 costs $4 and $20. Both charge $0.20 per million cached input tokens. Sol's price rises to $4/$15 for prompts over 272K tokens, while Opus 5.5 has no long-context premium. Both give 50% off through their Batch APIs.

Do Opus 5.5 and GPT-6 Sol have the same context window?

Nearly. Opus 5.5 has a 1M-token context window. GPT-6 Sol lists 1,050,000 tokens, with a 922,000-token maximum input. Both cap synchronous output at 128K tokens. Opus 5.5 can produce up to 300K output tokens on the Message Batches API with a beta header.

Can I use GPT-6 Sol in Claude Code or Opus 5.5 in Codex?

Not officially. Claude Code is built around Anthropic's models and Codex around OpenAI's. Third-party routers and open-source agents can run either model, but each vendor's published agent results use its own harness. Expect different results in a different agent.

Did OpenAI publish Terminal-Bench or SWE-bench scores for GPT-6 Sol?

Not at launch. OpenAI's reported Sol numbers cover OSWorld 2.0 (60.5%), Agents' Last Exam (56.4%), AutomationBench 1.0.6 (33.2%) and DeepSWE 1.1 (68.8%). Terminal-Bench 4.0 and SWE-bench Pro figures in circulation refer to GPT-5.6 Sol, not GPT-6 Sol.

Is GPT-6 Luna a replacement for Opus 5.5?

No. Luna scores 37 on the AA Intelligence Index to Opus 5.5's 58. It is built for cheap, high-volume, well-scoped work like extraction and summarising. It works best as a routing layer beneath Sol or Opus 5.5.

Sources

• Anthropic: Claude Opus 5.5 announcement, model page, pricing, system card
• OpenAI: GPT-6 Sol model page, GPT-6 Luna model page, API pricing, launch post
• Artificial Analysis: Opus 5.5, GPT-6 Sol
• Coverage: VentureBeat, TechCrunch, 9to5Mac, OfficeChai

If your team is building products on Opus 5.5 or GPT-6 Sol and needs engineers who already work this way, Codersera can help you hire vetted remote developers who are fluent in agent-driven workflows.