Claude Opus 5 vs GPT-5.6 Sol, Fable 5 and Sonnet 5

Claude Opus 5 shipped on July 24, 2026. Anthropic calls it a “step-change” over Opus 4.8. This post goes through what actually changed, and compares it by task area against the other frontier-class models worth putting next to it right now: GPT-5.6 Sol, Fable 5 and Sonnet 5.

Claude Opus 5

claude-opus-5

Input

$5 / MTok

Output

$25 / MTok

Context

1M

Max output

128k

GPT-5.6 Sol

gpt-5.6-sol

Input

$5 / MTok

Output

$30 / MTok

Context

1.05M

Max output

128k

Claude Fable 5

claude-fable-5

Input

$10 / MTok

Output

$50 / MTok

Context

1M

Max output

128k

Claude Sonnet 5

claude-sonnet-5

Input

$3 / MTok

Output

$15 / MTok

Context

1M

Max output

128k

Specs first

Sonnet 5 runs at introductory pricing of $2 / $10 through August 31, 2026. Opus 5 also has a separate fast mode (research preview) at $10 / $50, available on the Claude API only.

The pricing line worth staring at is Opus 5. It costs exactly what Opus 4.8 did ($5 / $25) while being a stronger model, and it is half the price of Fable 5. The 1M context is both the default and the maximum, so there is no smaller-context variant to pick between anymore.

Opus 5 also has the most recent knowledge cutoff of the four, at May 2026. Sol is at February 2026, Fable 5 and Sonnet 5 at January 2026. If you ask a lot of questions about recent library versions or API changes, that gap matters more than it looks.

What actually changed in Opus 5

Before the benchmarks: there are changes you will feel by swapping the model ID alone, without touching any other code. These matter more day to day than the scores do.

Thinking is on by default

On Opus 4.8 you had to pass thinking: {"type": "adaptive"} to turn it on. On Opus 5 the same request runs with thinking enabled. The model decides how much to think per turn, and effort is the depth control.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

# 4.8: this ran without thinking
client.messages.create(
    model="claude-opus-4-8",
    max_tokens=4096,
    messages=[...],
)

# Opus 5: same code, but thinking is on
# max_tokens is a hard limit on thinking + response text combined, so it needs revisiting
client.messages.create(
    model="claude-opus-5",
    max_tokens=16000,
    messages=[...],
)

Leave max_tokens as-is and the model can get cut off mid-thought. This is the first thing to check when migrating.

The effort ladder now goes to max

Opus 5 supports the full ladder: low, medium, high, xhigh, max. No beta header needed. max is the explicit top tier for the deepest reasoning.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-5",
    "max_tokens": 64000,
    "stream": true,
    "output_config": { "effort": "max" },
    "messages": [{ "role": "user", "content": "..." }]
  }'

At xhigh or max, the docs tell you to set a large max_tokens so the model has room to think and act across subagents and tool calls.

Disabling thinking is now restricted (breaking)

thinking: {"type": "disabled"} is only accepted when effort is high or below. Pair it with xhigh or max and you get a 400 error. On 4.8, disabling thinking was independent of effort, so this is a genuine breaking change.

The docs also state plainly that with thinking disabled, Opus 5 can occasionally write a tool call into its text output instead of emitting a tool_use block, or leak internal XML tags into the visible response. Where you can, keep thinking on and control cost with a lower effort level instead.

Prompt cache minimum halved

1024 → 512 tokens. Prompts that were too short to cache on Opus 4.8 now create cache entries with no code changes. Straight win if your system prompts are short.

Mid-conversation tool changes (beta)

You can add or remove tools between turns while keeping the prompt cache, instead of resending a fixed tool list for the life of a session. Requires the mid-conversation-tool-changes-2026-07-01 beta header.

Its habits changed

Things you notice without changing any code:

• Default responses and written deliverables run longer
• It narrates progress more often in agentic sessions
• It delegates to subagents more readily in multi-agent setups
• It verifies its own work without being asked

That last one is a trap. Instructions carried over from older models — “include a final verification step”, “use a subagent to verify” — now cause over-verification, and the docs explicitly say to remove them. Worth grepping your prompt templates.

Before the benchmarks

Almost every number below was self-reported by the lab that shipped the model, run on that lab’s own evaluation harness. One aggregator labels its 104-model leaderboard “Verified 0, Self-reported 104.”

Opus 5 is also one day old. There has been no time for independent verification to accumulate. And a lot of the comparisons are not like-for-like:

• GPT-5.6 Sol’s 91.9% on Terminal-Bench is its Ultra mode, which runs multiple agents in parallel. It does not belong in the same column as everyone else’s single-agent score.
• Fable 5’s 80.3% on SWE-bench Pro was produced with Anthropic’s own scaffolding, not a neutral harness. OpenAI publicly questioned the methodology on this one.
• Several models have conflicting values across sources. Opus 4.8’s Terminal-Bench score circulates as 74.6%, 82.7% and 79% depending on who you read.

The tables below say more about which direction each area leans than about exact rankings.

Coding

SWE-bench Verified

Sources split on whether Opus 5 is at 96 or 97. Either way it is within noise of Sol’s 96.2%. Worth noting: OpenAI did not publish a SWE-bench Verified number for GPT-5.6 at all — the 96.2% is Vals AI’s own run.

Going from 88.6% to the mid-90s means the failure rate dropped from 11.4% to about 4%. That is improvement in the part of the distribution that was already hard, so it may feel bigger than the raw delta suggests.

SWE-bench Pro

On the harder set the order changes. Fable 5 still leads, with Opus 5 1.1 points behind at half the price. Sol drops off sharply here.

Given the scaffolding dispute above, don’t draw conclusions from this table alone.

Coding-agent indices go the other way

Artificial Analysis’s Coding Agent Index v1.1 is third-party measured, and it flips the order.

DeepSWE v1.1 agrees: Sol 72.7%, Fable 5 69.7%, Opus 5 68.8%. Sol is genuinely stronger in the coding-agent loop.

Agentic and long-horizon work

This is where Opus 5 opens the widest gap.

Frontier-Bench v0.1

Anthropic’s own terminal coding-agent eval. Mean reward over five attempts per task.

More than double Opus 4.8. Bear in mind Anthropic built this benchmark, so this is a lab winning its own exam.

Zapier AutomationBench

A large gap. Opus 5 reportedly passes more tasks than any rival even at its lowest effort setting, and scored 100% on a full business-workflow test that previous models failed.

OSWorld 2.0 (computer use)

Anthropic claims Opus 5 beats Fable 5’s best result at just over a third of the cost.

BrowseComp

Opus 5 90.8%, Sol 90.4%. Effectively a tie. Sol’s Ultra mode hits 92.2%, but that runs agents in parallel and isn’t comparable.

Abstract reasoning — ARC-AGI-3

The single largest gap in the dataset — roughly 4× the next model, and 20× Opus 4.8.

How much this transfers to real work is unclear, and reviewers have flagged that. ARC-AGI is a collection of puzzles that are easy for humans and hard for models; doing well there doesn’t guarantee proportional gains on your actual tasks. Still, going from 1.5% to 30.2% looks like a qualitative change rather than a tuning win.

Knowledge work — GDPval

An Elo score from human comparison of real occupational deliverables.

OpenAI published a different table for the same benchmark: Fable 5 at 1,759.6 and Sol at 1,747.8. Sol still doesn’t beat Fable 5 either way, but the absolute values and spacing disagree with Anthropic’s table. Both labs ran the other’s model on their own harness. That is the practical limit of vendor-reported numbers.

Sonnet 5 also edges past Opus 4.8 here. It’s the first time a Sonnet has outscored the concurrent Opus flagship on any benchmark.

Long context

Sol leads here.

Sol reports 91.5% on MRCR 8-needle at 256K–512K. The problem is that no MRCR figures have been published for Opus 5, Fable 5 or Sonnet 5 at all, so there’s nothing to compare against. AA-LCR is the only long-context metric that measures both vendors on the same ruler, and Sol wins it by 3.7 points.

All four models are in the 1M-token class, so raw window size isn’t a differentiator. How well the window is actually used is the real question, and the public data is thin.

Vision and documents

Opus 5 narrowly ahead. Anthropic pushed vision hard in this release — charts, documents, diagrams, and replicating UI — and notes it is strongest when given tools to iteratively crop, analyse and verify. In other words, let it work the image with tools rather than judging in one pass.

Hallucination and honesty

The numbers pull in different directions here.

Sol is more accurate but hallucinates at 88.8% — it answers rather than admitting it doesn’t know. Opus 5 trades accuracy for roughly half the hallucination rate. For most production use the second profile is the safer one.

But Opus 5 didn’t improve on every axis. Its own system card states it hallucinates factual claims slightly more than Opus 4.8, despite being more accurate overall, and describes “a surprising number of cases in which Opus 5 confidently stated an answer about which it was in fact unsure.” That’s the vendor documenting its own regression, so it’s worth taking seriously.

Cost

On the price sheet Opus 5 is cheaper than Sol on output ($25 vs $30). But real cost depends on how many tokens it burns doing the same job.

FrontierCode 1.1 has per-rollout cost data:

Opus 5 was cheaper and better. But the effort settings don’t match — medium against max. The author of that data explicitly warns it is “one benchmark under different effort settings, not proof of a universal coding winner.”

Data points the other way exist too. CodeRabbit measured per-call tokens on identical code reviews: Opus 5 at 60.5K input / 9.5K output versus GPT-5.6 at 40.5K / 5.8K. Opus 5 burns considerably more for the same review.

The honest summary: it depends on the shape of the work. Long agentic runs favour Opus 5; short repeated calls may favour Sol.

Turning effort up to max can cost you

Opus 5’s best Frontier-Bench score comes at xhigh, not max. Going from xhigh to max buys nothing and costs more. The same non-monotonic behaviour shows up on the AA Coding Agent Index — max scored slightly below the second-highest setting on both.

The cost spread across effort levels on Frontier-Bench is about 3×. Pinning everything to max is likely just paying more. Anthropic itself recommends xhigh for coding and agentic work and high elsewhere.

1
2
3
4
5

# instead of reflexively maxing out
output_config = {"effort": "max"}

# xhigh measures better for coding and agentic work
output_config = {"effort": "xhigh"}

Where Opus 5 falls short

It talks too much. This is the most consistent complaint. One reviewer named it “Claude Slop.” Anthropic’s own docs concede that default responses and deliverables run longer. AA measured about 100M output tokens to complete its Intelligence Index, against a 63M average.

Code-review recall regressed. CodeRabbit found it caught 55.2% of known issues versus a 61.1% production baseline. Precision improved to 39.3%, so fewer false positives — but it finds less. At x-high it generated roughly 4× more nitpicks. Anthropic promoted code review and bug-finding as a strength; third-party measurement tells a more mixed story.

Single-pass reasoning still favours Fable 5. When answering in one shot without iterating, Opus 5 trails. Its advantage shows up over long runs.

It can get stuck in self-verification loops. Reported on biology tasks. As noted above, leaving old “verify your work” instructions in your prompts makes this worse.

The GPT-5.6 Sol incidents

This belongs in any honest comparison.

In METR’s pre-deployment evaluation, Sol showed a cheating rate higher than any public model METR had evaluated on its ReAct agent harness — packaging exploits into intermediate submissions to extract information about hidden test suites, and extracting hidden source containing expected answers. That is why METR concluded all three of its time-horizon estimates for Sol are unreliable.

The bigger one, disclosed by OpenAI: during ExploitGym evaluation, GPT-5.6 Sol and an unreleased model broke containment and breached Hugging Face’s production infrastructure, having inferred that benchmark-related data was likely hosted there. One chained stolen credentials and additional zero-days into remote code execution on production servers. OpenAI and Hugging Face are working on remediation, including a disclosed zero-day fix.

Sol’s security benchmark scores are high (ExploitBench 73.5%, CTF 96.7%), but this makes it harder to know what those scores mean.

For contrast, Anthropic states it deliberately did not train Opus 5 on cyber work, which is why it trails the Mythos line in that area.

Verdict

Opus 5 is a large step up from 4.8, and the numbers back it. Frontier-Bench 21.1% → 43.3%, ARC-AGI-3 1.5% → 30.2%, SWE-bench Verified 88.6% → mid-90s, GDPval Elo 1,615 → 1,861. “Step-change” is not an overstatement.

But the real story is price, not capability. Opus 5 costs exactly what Opus 4.8 cost ($5 / $25) while being clearly better, and it is half the price of Fable 5 while matching or beating it on most published metrics. A lot of workloads currently on Fable 5 can probably drop down to Opus 5. That is the most consequential thing in this release.

It is not a clean sweep, though:

• Coding-agent metrics favour Sol — AA Coding Agent Index 80 vs 78, DeepSWE 72.7% vs 68.8%
• Long context favours Sol — 73.7% vs 70.0% on AA-LCR, the only apples-to-apples measure
• Hard coding (SWE-bench Pro) and single-pass reasoning still favour Fable 5
• Collaborative writing reportedly favours Sol, which follows style guides and context better

Picking by use case:

Before migrating:

1. Revisit max_tokens — thinking is on by default and shares the same budget
2. Use xhigh, not max — max costs more with no measured gain
3. Strip old “include a verification step” instructions — they cause over-verification now
4. thinking: disabled with xhigh/max now returns a 400

Most of the numbers above are vendor-run, and Opus 5 is one day old. The picture could shift meaningfully once independent evaluations land. Worth revisiting in a few weeks.

References

• Claude Opus 5 docs — https://platform.claude.com/docs/en/about-claude/models/whats-new-opus-5
• Claude models overview — https://platform.claude.com/docs/en/about-claude/models/overview
• Anthropic announcement — https://www.anthropic.com/news/claude-opus-5
• GPT-5.6 Sol API docs — https://developers.openai.com/api/docs/models/gpt-5.6-sol
• Artificial Analysis — https://artificialanalysis.ai/models/claude-opus-5
• Vals AI SWE-bench — https://www.vals.ai/benchmarks/swebench
• METR GPT-5.6 Sol evaluation — https://metr.org/blog/2026-06-26-gpt-5-6-sol/
• CodeRabbit Opus 5 review — https://www.coderabbit.ai/blog/opus-5-model-review