Claude Opus 5 In-Depth ReviewWhat Really Changed vs Opus 4.8, Fable 5, and GPT-5.6 Sol

Bottom Line: Opus 5 Is Not a Cheaper Fable, but the New Everyday Workhorse

On July 24, 2026, Anthropic released Claude Opus 5. The model is available across Claude's paid plans, Claude Code, and the Claude API. API pricing is $5 per million input tokens and $25 per million output tokens: unchanged from Opus 4.8 and exactly half the list price of Fable 5.

The short version is that Opus 5 is not a small point release disguised by a major-version number. It makes substantial gains in real code repair, computer use, business automation, long-horizon knowledge work, and unfamiliar problem solving. Anthropic's official system card shows FrontierBench rising from 21.1% on Opus 4.8 to 43.3%, SWE-bench Multimodal from 38.4% to 59.4%, AutomationBench from 17.0% to 26.0%, and ARC-AGI-3 from 1.5% to 30.2%. The sticker price is unchanged, but the range of work that can be delegated is clearly different.

The relationship with Fable 5 is even more surprising. Fable still leads on SWE-bench Pro, DeepSWE, and a healthcare evaluation, but Opus 5 leads on FrontierBench, OSWorld 2.0, GDPval-AA, AA-Briefcase, and AutomationBench. So the answer to "does Opus 5 still fall short of Fable 5?" is: it still falls short on parts of the highest ceiling, but it has already moved ahead on many everyday agentic workloads.

There is no one-directional verdict against GPT-5.6 Sol either. Opus 5 leads on FrontierBench, FrontierCode, OSWorld 2.0, GDPval-AA, AA-Briefcase, and ARC-AGI-3, while Sol leads on DeepSWE and ARC-AGI-2. Model capability no longer collapses neatly into one score. The type of work, the agent harness, and the amount of test-time compute increasingly decide the winner.

What Changed: Same Price, Thinking on by Default

The API model ID is claude-opus-5. It has a default one-million-token context window, a 128,000-token maximum output, and a May 2026 reliable-knowledge and training-data cutoff. It is available not only through the first-party Claude API, but also through Amazon Bedrock, Google Cloud, and Microsoft Foundry.

The most consequential API change is that adaptive thinking is enabled by default. Opus 4.8 ran without thinking unless the caller explicitly set thinking: {"type": "adaptive"}. Opus 5 decides on each turn whether and how deeply it should think. The effort parameter has five levels — low, medium, high, xhigh, and max — with high as the default in the Claude API and Claude Code.

Other changes include a beta capability to add or remove tools between turns while preserving the prompt cache, a fallbacks: "default" mode that delegates recommended fallback selection to Anthropic, and a lower minimum cacheable prompt length of 512 tokens instead of 1,024 on Opus 4.8. Fast mode is available only through the Claude API and trades roughly 2.5 times the throughput for twice the standard price: $10 input and $50 output per million tokens.

Benchmark Overview: How Much More Accurate Is It Than Opus 4.8?

The following values come from the capability summary table in Anthropic's Claude Opus 5 System Card. Unless otherwise noted, Opus 5 used adaptive thinking at max effort and results were averaged over five trials. Bold marks the leader among the four displayed models.

The important point is that these are not uniform two- or three-point gains. FrontierBench measures whether a coding agent can carry real software tasks through to completion, and Opus 5 reaches about 2.05 times the Opus 4.8 resolution rate. SWE-bench Multimodal, which adds screenshots and visual specifications to code issues, improves by 21 points. AutomationBench tests end-to-end business workflows spanning systems such as CRM, Slack, and Google Workspace; there the relative improvement over Opus 4.8 is roughly 53%.

ARC-AGI-3 is the most dramatic result. Models enter unfamiliar interactive game environments without rules or stated goals and are scored against human action efficiency. Opus 5 reaches 30.16%, compared with 1.52% for Opus 4.8 and 7.78% for GPT-5.6 Sol — about 3.9 times Sol's score. There are important caveats: ARC-AGI-3 is new and semi-private, Opus 5 used high effort, Sol used max, and Opus 4.8 used high, so the settings are not uniform. It is a remarkable gap, but it does not mean "general intelligence is four times higher."

The Practical Difference from Opus 4.8: More Verification, Not Just Better Generation

The most meaningful Opus 5 improvement is less about writing a beautiful answer or clean code on the first pass and more about noticing what is missing, creating the necessary tools, verifying the result, and only then declaring the job complete.

In one FrontierBench example described by Anthropic, the model was asked to reconstruct a machine part as a FreeCAD model from a drawing, but the environment intentionally gave it no direct way to view the image. Rather than stopping, Opus 5 wrote its own computer-vision pipeline to extract geometry from raw pixels and reconstructed the part. It succeeded repeatedly; competing models in the same setup failed after five attempts.

In another case, it addressed a real package-manager bug and fixed an edge case that the community's existing patch had missed. While building a market-data feed, it discovered there was no live feed against which to validate the parser, so it built its own test harness. The significant behavior is not simply having more knowledge. It is the stronger tendency to create a verification method when direct evidence is unavailable.

Early customer evaluations point in the same direction. Box reported an 8% overall improvement over Opus 4.8, including 11% in data analysis and 17% in due-diligence workflows. One financial-modeling evaluation reported nine percentage points higher accuracy on average, one-third fewer turns and tool calls, and 60% less time. A Legal Agent Benchmark user reported comparable performance to Opus 4.8 at max reasoning while generating 26% fewer tokens on average. These figures appeared as customer reports on Anthropic's launch page, however; they are not fully independent reproduction studies.

Does It Reach Fable 5? Where Opus Has Passed It, and Where It Has Not

It is reasonable to ask why Opus 5 exists when Anthropic already offers Fable 5. Fable costs $10 per million input tokens and $50 per million output tokens and remains positioned as Anthropic's most capable widely released, Mythos-class model. Yet in the published table, Opus 5 wins more of the listed evaluations.

For routine computer use, enterprise artifacts, business automation, and general agentic coding, Opus 5 either exceeds Fable 5 at half the token price or reaches near parity. Anthropic says Opus 5 surpassed Fable 5's best OSWorld 2.0 result at a little over one-third of the cost, and came within 0.5 percentage points of Fable 5's peak on CursorBench 3.2 at max effort.

That does not make Fable obsolete. It retains a narrow SWE-bench Pro lead and performs better on DeepSWE and the listed healthcare benchmark. Early users have also reported that Fable is less prone to premature conclusions in ambiguous root-cause investigations and can hold a complex line of reasoning more reliably. Anthropic's own selection guidance still says to start with Opus 5 for complex agentic coding and enterprise work, while using Fable 5 when the highest available capability is required.

The practical strategy is therefore to make Opus 5 the default and escalate to Fable only when defining the problem is itself difficult, the cost of a wrong conclusion is extreme, or specialized medical and scientific reasoning demands the higher ceiling. The old boundary — Opus 4.8 for daily work, Fable for hard work — has moved substantially toward the Fable end.

Against GPT-5.6 Sol: Where Claude Leads and Where Codex Still Matters

GPT-5.6 Sol costs the same $5 per million input tokens, with output at $30, and has a 1.05-million-token context window. The price band is almost identical, but the official comparison table shows different strengths.

Read in isolation, this looks like an Opus landslide. The comparison still needs care. Some competitor figures were taken from developer publications or benchmark leaderboards rather than a single Anthropic-run harness under perfectly uniform settings. GPT-5.6 Sol also changes character when combined with Codex, its reasoning controls including Ultra, and parallel multi-agent execution. Existing GPT-5.6 evaluations have shown strengths in terminal work, browsing, computer use, scientific tasks, and producing useful artifacts with relatively compact outputs.

In one sentence: Opus 5 is especially attractive when requirements are ambiguous and the model must clarify the specification, move across documents, interfaces, and code, verify its work, and deliver a finished artifact. GPT-5.6 Sol remains compelling when a clear engineering task can be decomposed inside the Codex tool environment, multiple investigations or implementations can run in parallel, and short iteration loops matter. Real evaluation should include output tokens, retries, tool calls, and human review time — not just API list prices.

Day-to-Day Use: Existing Prompts Work, but Remove Legacy Over-Instructions

Anthropic says prompts written for Opus 4.8 generally work with Opus 5. But instructions added to compensate for older models — "always perform a final verification step" or "use another agent to check your work" — may now be counterproductive. Opus 5 verifies proactively, so stacking the same instruction can produce over-verification and waste both time and tokens.

What should be explicit is scope and desired length. Opus 5 readily narrates its progress during agentic work, and its per-message output is often longer than previous Opus models. Reports and documents written to disk also tend to become longer. Anthropic recommends controlling visible response length directly in the prompt rather than lowering effort. Effort controls how much the model thinks; it does not reliably control how much it says.

Another tendency is scope expansion at high effort. On FrontierCode, performance declined above high because the model sometimes made extra refactors or other improvements outside the requested change. The grader expected a minimal, mergeable diff. Anthropic found that a short instruction to stay within task scope recovered much of the loss. In a large repository, it is worth spelling out target files, prohibited changes, the allowed boundary, and completion criteria.

A practical starting policy is low or medium for daily work, high for difficult debugging and design, xhigh for high-cost one-off tasks, and max only for the final class of problems that truly needs it. Many headline results use max, but GDPval-AA xhigh still beat every other model while using 25% fewer output tokens than max, and AutomationBench medium scored 24% at $0.89 per task. Max effort is not automatically the best performance per dollar.

Launch-Day User Reports: Amazement and Frustration Arrived Together

The model has only been public for about a day, so user reports are anecdotes, not statistics. They are still useful for understanding the model's emerging character.

Positive reactions include claims that Opus 5 feels clearly faster than Opus 4.8 and writes better code, that Low effort can outperform Sonnet 5 High or Opus 4.8 High, and that long-running sessions consume less weekly allowance than expected. Another user compared fresh chats on the same complex issue and ranked Opus 5 first, followed by Fable 5, Opus 4.6, Sonnet 4.6, and Opus 4.8. It was one private test, but the idea that Opus 5 Low may be the new cost-performance sweet spot is consistent with Anthropic's effort curves.

The negative reports matter just as much. A developer who has worked for more than a year on the same large project across multiple models said Opus 5 produced sound code but made hasty root-cause conclusions, treated a comment as evidence, and repeatedly reversed its claim about a default flag. Auditing the reasoning with Fable allegedly revealed several additional issues. For ambiguous investigation, that user still trusted Fable more. In a separate API test, one user reported that matched-effort Opus 5 produced roughly twice as many output tokens as Opus 4.8 and hit a 32k cap at medium and high. That test was only one run on one task type and had not yet scored output quality, so it should not be generalized.

CodeRabbit ran three trials across roughly 100 error patterns drawn from verified real open-source pull requests. Opus 5 at xhigh produced a highly precise actionable stream but lower coverage, while the broader output still contained noise. The company judged it a stronger builder than reviewer, with better design judgment than Opus 4.8, but slower and less efficient than Fable 5 and overly cautious around security. The shared lesson is that Opus 5 is not universally best at every setting: precision, coverage, length, and cost move with the task and effort level.

Safety and Limits: Anthropic's Most Aligned Model, but Deliberately Not the Cyber Frontier

In Anthropic's automated behavioral audit, Opus 5 scored 2.3 on overall misaligned behavior, lower than Opus 4.8, Sonnet 5, and Fable 5. Anthropic also positions it as the company's strongest model so far at avoiding deceptive behavior, susceptibility to misuse, and reckless actions with hard-to-reverse consequences.

There is an intentional capability ceiling in sensitive domains. As with Opus 4.8, Anthropic says it did not train Opus 5 specifically on cyber tasks. Stronger general intelligence brings vulnerability discovery close to Mythos 5, but exploit development — converting a vulnerability into a material threat — remains substantially behind Mythos. It also does not advance the frontier in biology research or offensive cybersecurity.

For most developers, that is not necessarily a drawback. Opus 5 can deliver high general capability without the special operating constraints associated with Fable 5, such as its 30-day retention requirement and broader fallback design. Legitimate security researchers may still encounter refusals or excessive caution, however. The intended division is Mythos 5, by invitation, for the highest-end approved defensive cyber work, and Opus 5 for ordinary code review and safe remediation.

Capability limits remain visible elsewhere too. A 79.2% SWE-bench Pro score means roughly one in five tasks still fails. AutomationBench is state of the art at 26%, yet that means about three-quarters of complex business workflows are not completed. OSWorld 2.0 is 70.6%, not 100%. A model that verifies its own work does not eliminate the need for human permission design, diff review, tests, and explicit approval for external communications or production changes.

Which Model Should You Choose? A Practical July 2026 Guide

For individual Claude users, Opus 5 can now be the default. The main reasons to remain on Opus 4.8 are regression control in an existing production system or not yet having tuned Opus 5's tendency toward longer outputs. API migration is not only a model-ID change: thinking is now on by default, manual extended thinking and some sampling configurations are incompatible, and both conversational answers and written artifacts may become longer.

Fable 5 is no longer simply "the premium option that always gives the best answer." Its higher cost is easiest to justify when the problem definition is unclear, premature reasoning would create serious losses, and Opus 5 has already reached a genuine impasse. GPT-5.6 Sol likewise should be judged as part of Codex's execution environment, not only as a row in a model table.

After deployment, build at least several dozen representative internal tasks and record completion rate, number of corrections, elapsed time, output tokens, tool calls, human review time, and serious misses. Raising effort does not guarantee a better result. The same model can behave like a different product at low and max because its cost and scope tendencies change so much.

Conclusion: More Than a Model Upgrade, This Is a Shift Toward the Verifying Agent

The essence of Claude Opus 5 is not a few more benchmark points. If it cannot see an image, it may build a tool to inspect the pixels. If validation data is missing, it may create a test harness. If it builds an interface, it can inspect desktop and mobile views. If it finds an error, it is more likely to correct it before handoff. The product is moving from a model that generates answers toward an agent that keeps working until completion criteria are satisfied.

Against Opus 4.8, this is a clear generational replacement at the same list price. Against Fable 5, the half-price Opus model has taken over as the practical leader on most everyday work while leaving Fable a narrower set of ceiling cases. Against GPT-5.6 Sol, Opus 5 looks stronger on novel problems, computer use, and professional knowledge artifacts, while Sol retains advantages in long-horizon coding and the parallel Codex execution environment.

As of July 25, 2026, the practical conclusion is straightforward: if you use Claude for work, Opus 5 can be your new baseline. Fable 5 becomes an escalation path for the hardest cases, GPT-5.6 Sol remains a different but powerful execution system, and Opus 4.8 becomes the compatibility baseline you are migrating away from. Launch-day user reports are still preliminary. Stability over weeks, speed under load, real usage consumption, and reasoning quality deep into long sessions all deserve continued measurement.

References

1. Introducing Claude Opus 5 — Anthropic
2. Claude Opus 5 System Card — Anthropic
3. What’s new in Claude Opus 5 — Claude Platform Docs
4. Models overview — Claude Platform Docs
5. Prompting Claude Opus 5 — Claude Platform Docs
6. Introducing Claude Fable 5 and Claude Mythos 5 — Anthropic
7. GPT-5.6 Sol — OpenAI
8. Claude Opus 5 provider performance — Artificial Analysis
9. Claude Opus 5 benchmarks for AI code review — CodeRabbit
10. Opus 5 first impressions vs Fable — Reddit
11. Claude Opus 5 early efficiency discussion — Reddit
12. Claude Opus 5 vs Opus 4.8 token-use test — Reddit