Claude Opus 5 vs Grok 4.5 vs GPT-5.6 Sol: What Changed for Builders

Anthropic released Claude Opus 5 on July 24, 2026. It's the third frontier coding model in about two weeks: xAI and Cursor shipped Grok 4.5 on July 8, and OpenAI opened GPT-5.6 to everyone on July 9. Opus 5 is one day old as I write this, so what follows is the launch data, the first independent tests, and the migration notes — not our own benchmark run. We'll publish that once we've put it on real client workloads.

The headline is price. Opus 5 keeps Opus 4.8's rate of $5 per million input tokens and $25 per million output, and Anthropic's own numbers put it close to Claude Fable 5, which costs $10/$50. If those claims survive contact with your workload, the cost per task on agentic work drops by roughly half.

What Claude Opus 5 is, in numbers

• Model ID claude-opus-5. $5 per million input tokens, $25 per million output.
• 1 million token context window — that's both the default and the maximum, with no long-context surcharge. Up to 128K output tokens per request.
• Five effort levels (low, medium, high, xhigh, max), defaulting to high. Effort is the main cost and latency dial.
• Thinking is on by default, which is a change from Opus 4.8 where omitting the parameter meant no thinking. Disabling it is only allowed at high effort or below.
• Fast mode runs the same model at up to 2.5x the output speed for $10/$50. Claude API only — not on Bedrock, Vertex or Foundry.
• The minimum cacheable prompt drops to 512 tokens, from 1,024 on Opus 4.8. Prompts you'd written off as too short to cache now cache.

Availability was broad on day one: Claude.ai, the API, Claude Code and Claude Cowork, plus GitHub Copilot on Pro+, Max, Business and Enterprise (billed at API list price; Business and Enterprise admins have to enable it), and Kiro at a 2.2x credit multiplier with the full 1M context in us-east-1 and eu-central-1.

How the three models compare on price

• Claude Opus 5: $5 in / $25 out per million tokens. 1M context.
• GPT-5.6 Sol: $5 in / $30 out. 1M context, February 2026 knowledge cutoff. Terra and Luna sit below it at lower rates.
• Grok 4.5: $2 in / $6 out, or $4/$18 for the fast variant. 500K context, down from Grok 4.3's 1M. Not available in the EU under the AI Act.
• Claude Fable 5, for reference: $10 in / $50 out.

Grok 4.5 is the cheapest by a wide margin, and Cursor's own measurements back that up in practice rather than just on the rate card: 66.7% on CursorBench 3.2 at an average of 33 steps and about $1.51 per task. Cursor also flags that an earlier snapshot of their codebase ended up in Grok's training data, which inflates its CursorBench score — worth remembering before you treat that particular number as a like-for-like.

What Anthropic claims, and how much of it to believe

The launch numbers are strong and, as always, vendor-selected:

• Frontier-Bench v0.1, a new 74-task successor to Terminal-Bench: 43.3% at max effort against 33.7% for Fable 5. State of the art, on a benchmark Anthropic introduced.
• CursorBench 3.2: within half a point of Fable 5 at half the price.
• OSWorld 2.0 computer use: ahead of Fable 5 at roughly a third of the cost.
• ARC-AGI 3: about three times the next-best model.
• Zapier's AutomationBench: roughly 1.5x the pass rate of the next model.
• Organic chemistry up 10.2 points and protein prediction up 7.7 points over Opus 4.8, with around 26% fewer tokens in some workflows.

Two caveats. Frontier-Bench is Anthropic's own benchmark and is at version 0.1, so treat it as a claim rather than a result. And Anthropic is explicit that Opus 5 sits behind Mythos 5 on cyber exploitation and autonomous biology work by design — requests that trip the safety classifiers get routed to Opus 4.8, which is a fallback you should handle in code rather than discover in production.

What independent testers found

This is where the picture gets more useful, because the outside results disagree with each other.

• Epoch AI put Opus 5 at 159 on its capability index against 161 for Fable 5, with the two tied on software engineering. On that measure the gap is small enough to be noise, and the price gap isn't.
• Artificial Analysis ranked it first on agentic knowledge work, ahead of Fable 5 by close to 150 Elo while cutting cost per task by about 20%.
• CodeRabbit measured 39.3% precision on actionable code review comments against a 35.2% baseline — but four times as many nitpicks, which is real triage work for whoever reads the output.
• Cognition's Scott Wu described it as approaching Fable-level performance at half the cost on debugging. Zapier's Wade Foster said it topped their automation leaderboard without using more tokens.
• Claire Vo called it 'brilliant but annoying' and noted it declined to handle merge conflicts.

Read together: the capability jump over Opus 4.8 is real but not enormous on aggregate indices, and the economics are what actually moved. Several developers argued the single-digit index gains understate what they see day to day, which is the kind of claim only your own evaluation can settle.

Where Grok 4.5 and GPT-5.6 Sol stand

Grok 4.5 came out of a joint effort between xAI and Cursor, trained partly on Cursor's own developer-agent interaction data rather than only on public repositories. It posts 83.3% on Terminal-Bench 2.1, 78.0% on SWE-Bench Multilingual and 64.7% on SWE-Bench Pro, and sits fourth on Artificial Analysis's intelligence index while leading on agentic tool use. The consistent hands-on report is that it's noticeably faster than Opus or the latest GPT models even without fast mode, reliable over long sessions, and weak on frontend design — the UI it generates is serviceable and no better. At $2/$6 it's the obvious candidate for high-volume, non-visual work.

GPT-5.6 Sol is OpenAI's flagship of that release, scoring 80 on the Artificial Analysis Coding Agent Index, a couple of points above Fable 5, and it's about 54% more token-efficient on coding tasks. Every's hands-on review is the most instructive thing published about it: fast enough for genuine live collaboration, strong at following a style guide, persistent through long assignments — and prone to overbuilding, in one case producing a 12,900-line system where something much smaller would have done. It also came last on their writing benchmark at 56 out of 100, behind Claude Opus 4.8 and Sonnet 5. OpenAI shipped GPT-Live alongside it on July 8, a full-duplex voice model that replaces Advanced Voice Mode for paying ChatGPT tiers.

The migration cost is in your prompts, not your code

The API surface barely changes if you're coming from Opus 4.8 — swap the model ID and you're running. The work is in the prompts, and most of it is deletion. Anthropic's own guidance for Opus 5 says this plainly:

• Delete your verification instructions. Opus 5 verifies its own work unprompted, and telling it to 'double-check' or adding a final verification step causes over-verification that burns tokens with no quality gain. This inverts a prompting rule most teams have been following for two years.
• Cap subagent delegation. Opus 5 delegates more readily than Opus 4.8, which under-delegated. Any 'use subagents more' guidance you added for 4.8 should come out, and a hard spawn limit should go in.
• Ask for brevity explicitly. Responses and written files both run longer than on prior models, and lowering effort does not reliably shorten visible output — only prompting does.
• Re-run an effort sweep. Low and medium are unusually strong here, and effort defaults carried over from another model are rarely right.
• In code review prompts, drop 'only report high-severity issues'. Opus 5 follows that literally and reports less; ask for everything and filter in a second pass.
• Don't disable thinking. With thinking off, the model occasionally writes a tool call into its visible text instead of emitting a real tool call — the turn succeeds, the call never runs, and nothing errors. Use low effort instead.

That last one matters most for anyone running agents. A tool call that silently doesn't happen is the worst failure mode in the list, because your harness sees a clean turn.

Which one to reach for

For multi-file features, larger refactors and long autonomous runs, Opus 5 is the strongest of the three on current evidence, and the price makes it defensible where Fable 5 wasn't. For high-volume agentic work where cost per call decides whether the feature is viable, Grok 4.5 at $2/$6 is hard to argue with as long as the work isn't visual and you're not shipping to the EU. For interactive, fast-iteration collaboration and spreadsheet-shaped analysis, Sol earns its place, with a human checking the scope of what it builds.

A leaderboard tells you how a model did on someone else's task. The only number worth acting on is how it performs on yours, measured against what you already run.

How we'd approach a switch

The same way we approach every model release: take one real task from a live project, run it against the incumbent and the challenger with the same harness, and count the things that matter — cost per completed task, how often a human has to intervene, and how many of the failures are recoverable. Most of our client systems end up using more than one model deliberately, because the cheapest model that clears the bar for a given call is almost never the same model everywhere.

If you're weighing an AI feature and want to know whether the new pricing changes what's affordable, that's a scoping conversation we're glad to have. Feasibility first, quote after.