The Vibe Father
Anthropic released Claude Opus 5 on July 24, 2026 at $5 per million input tokens and $25 per million output tokens — the same price as Opus 4.8, and half of Fable 5. The public model identifier is claude-opus-5, the context window is 1M tokens, and it is available on the Claude API, the Claude apps, Claude Code, and Claude Cowork.
The short version for people who ship software. Opus 5 is now the top model on the Artificial Analysis Intelligence Index, it takes the top spot on the Agentic Index, it ties for first on the Coding Index, and it does that at half the token price of the model it just passed. It is not a clean sweep. GPT-5.6 Sol still wins DeepSWE v1.1, Mythos 5 is still ahead on legal and on exploit development, and Fable 5 still leads on health tasks.
This article walks through every published number, separates independently measured results from vendor-reported ones, and ends with the routing decision that actually matters on Monday morning. If you want the head-to-head, read Opus 5 vs Fable 5 vs GPT-5.6 Sol. If you want the money math, read the pricing and effort-level guide.
The launch facts, without the marketing
Claude Opus 5 at a glance (July 24, 2026)
| Attribute | Value |
|---|---|
| Model ID | claude-opus-5 |
| Release date | July 24, 2026 |
| Input price | $5 per million tokens |
| Output price | $25 per million tokens |
| Context window | 1M tokens |
| Effort levels | low, medium, high, xhigh, max |
| Fast mode | ~2.5x default speed at 2x base rate |
| Surfaces | Claude API, Claude apps, Claude Code, Claude Cowork |
Two of those rows matter more than the rest. The first is price parity with Opus 4.8, this is a capability jump that costs the same as the model you are probably already paying for. The second is the effort ladder. Opus 5 exposes five reasoning-effort settings, and almost every headline benchmark result you will read today was measured at max effort. That is not the setting most production agents should run at, which we cover below.
Independently measured, the Artificial Analysis indices
Artificial Analysis publishes three capability indices on a fixed 0–100 scale, and they are the inputs to our own Vibe Coding Index. On launch day, all three were published for the adaptive-reasoning, max-effort configuration.
| Index | Claude Opus 5 | Next best published |
|---|---|---|
| Intelligence Index | 61 (#1) | Claude Fable 5 at 60, GPT-5.6 Sol (max) at 59 |
| Coding Index | 78 (joint #1) | GPT-5.6 Sol (xhigh) also 78, GPT-5.6 Sol (max) 77 |
| Agentic Index | 55 (#1) | GPT-5.6 Sol (max) at 54 |
Artificial Analysis capability boards, read July 24, 2026. Values are for the max-effort configuration.
Feed those three into our capability blend — 20% Intelligence, 45% Coding, 35% Agentic — and Opus 5 lands at a Vibe Coding Index of 66.6. That is a measured score built from three independent inputs, not a launch estimate, which is why the Opus 5 scorecard carries a real rank instead of a pending label.
Notice how narrow the coding margin is. A one-point gap on the Coding Index between Opus 5 and GPT-5.6 Sol (xhigh) — actually a tie — is not a reason to migrate a codebase. The gap that is genuinely large is on reasoning and on agentic knowledge work, and that is where the interesting numbers live.
Where the gap is actually big
Frontier-Bench v0.1 — agentic coding
Opus 5 at 43.3%, GPT-5.6 Sol at 34.4%, Fable 5 at 33.7%, Opus 4.8 at 18.7%. Anthropic's own framing is that Opus 5 doubles Opus 4.8 at a lower cost. Figures via Anthropic's announcement and Vellum's benchmark write-up.
Frontier-Bench is the result Anthropic leads with, and it is the one that most closely matches what an agent actually does, long-horizon coding work where the model has to keep going after the obvious approach fails. A 43.3% result against a 33.7% incumbent is a real jump, not a rounding difference.
The other outlier is ARC-AGI-3, which tests novel problem-solving on tasks the model has not seen a pattern for.
| Model | ARC-AGI-3 |
|---|---|
| Claude Opus 5 | 30.2% |
| GPT-5.6 Sol | 7.8% |
| Claude Opus 4.8 | 1.5% |
A near-4x lead over the next model is unusual enough to deserve skepticism, and it is worth saying plainly. ARC-AGI-3 is a research benchmark, not a proxy for whether the model can fix your flaky integration test. It tells you the model is better at problems with no memorized template. That correlates with debugging genuinely weird production bugs. It does not correlate with writing another CRUD endpoint, where a much cheaper model is fine.
The full published scoreboard
Every Claude Opus 5 figure with a public source, July 24, 2026
| Benchmark | Opus 5 | Comparison | Type |
|---|---|---|---|
| SWE-bench Verified | 96.0% | — | Reported |
| SWE-bench Pro | 79.2% | GPT-5.6 Sol 64.6% | Reported |
| SWE-bench Multilingual | 89.5% | — | Reported |
| DeepSWE v1.1 | 68.8% | GPT-5.6 Sol 72.7% | Reported — Opus 5 loses |
| Frontier-Bench v0.1 | 43.3% | Fable 5 33.7% | Reported |
| OSWorld 2.0 | 70.6% | GPT-5.6 Sol 62.6% | Reported |
| ARC-AGI-3 | 30.2% | GPT-5.6 Sol 7.8% | Reported |
| ARC-AGI-2 | 90.4% | — | Tracker |
| GPQA Diamond | 93.2% | — | Tracker |
| GDPval-AA v2 | 1,861 Elo | Fable 5 1,747, GPT-5.6 Sol 1,736 | Independent |
| AA-Briefcase | 1,720 Elo | +146 over Fable 5 | Independent |
| IMO 2026 | 42/42 | — | Tracker |
Read the "Type" column before you read the numbers. Independent means a third party ran the evaluation with its own harness. Reported means the figure comes from the launch materials or coverage of them, without a published run log you can reproduce. Tracker means an aggregator lists it, and aggregators disagree — SWE-bench Verified is quoted at 96.0% in one place and 97.0% in another, which usually means different trial counts or different scaffolds. Treat one-point differences on a reported benchmark as noise.
The DeepSWE row is the honest one. GPT-5.6 Sol wins it by roughly four points. Any launch article that lists twelve wins and no losses is selling something.
Knowledge work is the quiet headline
GDPval-AA v2 measures how well a model does real professional knowledge work through an open-source reference agent harness. Opus 5 scores 1,861 Elo there, more than 100 points clear of both Fable 5 and GPT-5.6 Sol. On AA-Briefcase, an agentic knowledge-work benchmark, it leads Fable 5 by 146 Elo.
This matters more than the coding indices for most teams, because most engineering time is not spent generating code. It is spent reading tickets, reconciling requirements, drafting migration plans, and writing the document that explains why the change is safe. Those are exactly the tasks a knowledge-work benchmark measures, and they are exactly the tasks where a weaker model produces confident output that a senior engineer then has to unwind.
Anthropic also reports domain gains outside software. 10.2 percentage points over Opus 4.8 on organic chemistry evaluations and 7.7 points on protein tasks. If you build in life sciences, those are worth your own private eval before you take them at face value.
Where Opus 5 is not the answer
Published weaknesses, stated plainly
| Area | What the evidence says | What to do |
|---|---|---|
| DeepSWE v1.1 | GPT-5.6 Sol leads 72.7% to 68.8% | Keep a Sol route for repo-scale SWE work |
| Exploit development | Substantially behind Mythos 5 | Do not expect offensive security capability |
| Legal tasks | Mythos 5 outperforms | Route legal analysis elsewhere |
| Health tasks | Fable 5 leads | Keep Fable 5 for clinical-adjacent work |
| Latency | ~62.7s to first token at max effort | Never put max effort in an interactive path |
The latency number deserves emphasis because it is the one most likely to surprise you. Artificial Analysis measures roughly 52.8 output tokens per second and about 62.7 seconds to first token for the max-effort configuration. That is a reasoning model thinking before it speaks. It is completely fine for a background agent working through a ticket queue. It is unusable for autocomplete, and it will feel broken in a chat box unless your UI streams a thinking state.
On security, Anthropic is specific. Opus 5 is close to Mythos 5 on identifying vulnerabilities in OSS-Fuzz and substantially behind on developing exploits. That gap is deliberate. Anthropic reports that cyber classifiers trigger about 85% less often than on Fable 5, with blocked requests falling back to Opus 4.8. In practice that means defensive security research works better than it did, and offensive tooling is still not the product you are being sold.
The effort ladder is the real product decision
Opus 5 exposes five effort levels. The published benchmark numbers are almost all from max. Reporting also notes something counterintuitive, max effort occasionally scores below xhigh on a couple of benchmarks while costing more. More thinking is not monotonically better.
A practical effort policy
| Effort | Use it for | Avoid it for |
|---|---|---|
| low | Formatting, renames, mechanical edits | Anything with a design decision in it |
| medium | Routine feature work with tests | Cross-service refactors |
| high | Most agent work worth paying for | Trivial edits — you are burning tokens |
| xhigh | Hard debugging, migration planning | Interactive latency-sensitive paths |
| max | Benchmarks, and genuinely stuck problems | Default production routing |
If you take one operational lesson from launch day, take this one, pick your effort level from your own evals, not from the press release. The vendor optimizes the number that makes the chart look good. You are optimizing accepted pull requests per dollar.
How this changes our leaderboard
Opus 5 enters our board with a measured Vibe Coding Index of 66.6, built from Intelligence 61, Coding 78, and Agentic 55. Because all three source dimensions are independently published, it gets a normal numbered rank rather than the tentative label we applied to Kimi K3 at its launch, when only part of the profile was available.
Two things will move that number in the next few weeks. First, the OpenRouter benchmark feed will list the model, at which point our live sync takes ownership of the scorecard and the hand-mirrored figures are deleted. Second, Arena Code WebDev human-preference votes will accumulate, and our coding dimension blends that signal at a fixed 35% weight once fresh votes exist. Expect the coding number to move by a point or two in either direction. That is the system working, not a correction.
What we would actually do with this
- Do not migrate anything today. Launch day is the worst day to change a default.
- Run your own suite, twenty real tickets from your backlog, scored on accepted diff, repair time, and token cost.
- Compare Opus 5 at high effort against whatever you run now. Not max — high. That is the honest production comparison.
- If Opus 5 wins on accepted work per dollar, change the default and pin the model revision.
- Keep a second route alive. GPT-5.6 Sol still wins DeepSWE, and single-vendor dependency is a business risk, not a technical preference.
The price parity with Opus 4.8 is what makes this an easy evaluation to justify. You are not being asked to double your inference budget to test a hypothesis. You are being asked to point an existing budget at a better model and see whether your review queue gets shorter.
Common questions
How much does Claude Opus 5 cost?
$5 per million input tokens and $25 per million output tokens — identical to Opus 4.8 and half of Fable 5's $10/$50. Fast mode runs about 2.5x faster at twice the base rate, which works out to roughly $10/$50.
Is Claude Opus 5 better than GPT-5.6 Sol for coding?
They tie on the Artificial Analysis Coding Index at 78. Opus 5 leads on SWE-bench Pro, OSWorld 2.0, and agentic knowledge work, GPT-5.6 Sol leads on DeepSWE v1.1. For most teams the deciding factor will be cost per accepted change on your own repository, not the leaderboard.
What is Claude Opus 5's context window?
1M tokens. That does not mean you should fill it — retrieval discipline still beats dumping a monorepo into the prompt.
Should I use max effort?
Usually not. Max is the benchmark setting. High or xhigh is the production setting for most agent work, and reporting notes max occasionally underperforms xhigh while costing more.
Where does Opus 5 rank on the Vibe Coding Index?
It enters at 66.6, built from published Intelligence (61), Coding (78), and Agentic (55) indices. See the full scorecard for the per-dimension sources.
Sources and further reading
- Anthropic. Introducing Claude Opus 5
- Artificial Analysis. Claude Opus 5 model profile
- Vellum. Claude Opus 5 benchmarks explained
- The Decoder, near-Fable 5 performance at half the price
- BenchLM. Claude Opus 5 benchmark tracker
- How to rank coding models without fooling yourself
One closing note on method. Every launch cycle produces a wave of articles that copy a vendor chart, add adjectives, and publish. The useful version of this work is boring, label which numbers were independently measured, name the benchmarks the model lost, and state the latency cost of the setting that produced the headline. If a model is genuinely good, it survives that treatment easily. Opus 5 does.
Keep a written record of what you changed and why. When the next model lands — and on the current cadence that will be weeks, not quarters — you want a one-page note that says which lane each model owns, what your fallback is, and which private eval would have to move before you touch the defaults again. Teams without that note re-argue the same decision every launch. Teams with it spend launch day reading, running twenty tickets, and getting back to work.
