Anthropic Opus 5: Half Fable 5's Price, Near Its Score

> **TL;DR:** Anthropic has released Opus 5, a model that costs roughly half as much as Fable 5 while performing close to it — on Cursor's hardest coding benchmark it reportedly lands within half a percent of Fable 5 at about half the cost per task. Fable 5 remains Anthropic's top-intelligence model, but the quality gap is now small enough that price, not capability, is the deciding factor for most workloads. Opus 5 is also available on the $20/month Claude Pro plan, moving capability previously associated with the $200 tier down the pricing ladder.

Key Takeaways

- Opus 5 costs roughly half of Fable 5 and performs close to it — the intelligence went up without the price going up with it. - On Cursor's hardest coding benchmark setting, Opus 5 reportedly scores within half a percent of Fable 5 at about half the cost per task. - In an Anthropic test, Opus 5 was given an image it could not view and wrote its own computer vision code to extract the shape and rebuild the part in 3D. - Opus 5 is available on the $20/month Claude Pro plan, not just premium tiers. - Fable 5 is still the smarter model — it just has to earn a 2× premium now instead of being the automatic default.

Anthropic has released **Opus 5**, and the most consequential number attached to it isn't a benchmark score — it's the price. Opus 5 costs roughly half of what Fable 5 costs while performing close to it. Fable 5 is still Anthropic's top-intelligence model. The question this release forces on every team with a model-selection config file is whether "still smarter" is worth twice the money.

For a lot of workloads, the answer just flipped.

The short version: the price-per-intelligence curve bent

Two facts define this launch. First, Opus 5 is a new Anthropic model priced at approximately half of Fable 5. Second, Fable 5 remains the more capable of the two. Stated that flatly it sounds like ordinary product tiering — big model, cheaper model, pick the one that fits your budget.

What makes it worth writing about is that the usual trade-off didn't show up. Moving down a tier has historically meant accepting a visible drop in reasoning quality: more retries, more hand-holding, more instructions that the smaller model quietly ignores. Opus 5 appears to break that pattern. Intelligence went up; the price didn't follow it. That's why we'd file this as one of the more underrated releases of the current cycle — it doesn't raise the ceiling, it raises the floor underneath the ceiling, which is the change that actually reaches production systems.

The Cursor benchmark result

The cleanest evidence so far comes from coding. On Cursor's coding benchmark at its most difficult setting, Opus 5 reportedly scores within half a percent of Fable 5 — while costing about half as much per task.

Half a percent is inside the range where most eval harnesses can't confidently distinguish two models. Read strictly, it means the hardest setting of a serious, industry-watched coding benchmark can no longer separate Anthropic's flagship from its half-price sibling in a way that would change your decision.

And cost per *task* is the metric that matters here, not cost per token. Agentic coding doesn't make one call — it makes dozens, sometimes hundreds, chained across file reads, edits, test runs, and retries. A 2× unit-price difference compounds across every step of that loop. At the scale where a team runs an agent continuously against a real codebase, halving cost per task isn't a line-item optimization; it's the difference between an experiment and a budget line.


The computer vision test says more than the benchmark does

The more interesting result is one Anthropic ran itself. The company gave Opus 5 an image of a machine part and asked it to rebuild the part as a 3D model — but gave it no means of actually viewing the image.

The expected outcomes here are failure or fabrication: refuse the task, or hallucinate a plausible-sounding part and present it confidently. Opus 5 did neither. It wrote its own computer vision code to extract the shape from the raw pixel data, then used that extracted geometry to reconstruct the part.

Anthropic reports that no other model completed this task, even after five attempts. That's a vendor-run evaluation, so we'd hold it loosely until independent parties reproduce it — but the shape of the behavior is what's notable, not the scoreboard. The model wasn't given a capability, so it built the capability, then used it. That's a step past tool use into tool *creation*.

Why "builds its own tools" is the actual headline

The last stretch of AI infrastructure work has been overwhelmingly about handing models tools in a standardized way. That's the whole premise behind [the Model Context Protocol ecosystem](https://speka.info/blog/model-context-protocol-the-repo-behind-every-mcp-tool), and it's why [Claude Skills repos have been climbing GitHub trending](https://speka.info/blog/claude-skills-repos-take-over-github-trending) — developers packaging capabilities so a model doesn't have to improvise them.

A model that writes its own tooling when the prepackaged version is missing doesn't make that infrastructure obsolete. Curated tools are still faster, more reliable, and more auditable than improvised ones. What it does change is the failure mode. Today, a gap in your tool coverage is a hard stop — the agent hits the missing capability and gives up. If the improvisation behavior generalizes, a coverage gap becomes a slowdown instead: the model routes around it by writing the missing piece. That's a meaningfully different reliability profile for anything running unattended.

The $20 tier is the part that moves the market

Opus 5 is available on the Claude Pro plan at $20 per month. Capability previously associated with the $200 tier now sits in the entry-level paid plan.

This is the change most likely to be felt outside developer circles. Pricing ladders in AI have generally been built on the assumption that frontier-adjacent capability is a premium good — you pay ten times more to get the good model. Compressing that gap resets the anchor for what $20 is supposed to buy, and every competing subscription tier now has to answer for the difference. It also broadens who gets to build seriously: an independent developer working through [Claude Cookbook patterns](https://speka.info/blog/claude-cookbook) on a Pro plan is now running close to what an enterprise seat ran on a few months ago.


When Fable 5 is still the right call

None of this makes Fable 5 obsolete. It remains the smarter model, and "within half a percent on one benchmark" is not "identical everywhere." Reach for Fable 5 when:

- **The task is long-horizon and self-correcting is expensive.** Small reasoning-quality differences compound across very long agent runs. If a wrong turn on step 40 costs you the whole run, buy the better reasoner. - **You're at the edge of the benchmark, not the middle of it.** Aggregate scores hide tails. Novel, ambiguous, or genuinely hard problems are exactly where a half-percent aggregate gap can widen. - **The cost is trivially small relative to the stakes.** For low-volume, high-consequence work — legal review, architecture decisions, anything a human will act on directly — a 2× price on a handful of calls is not a real constraint.

For high-volume, well-scoped, repeatable work — bulk classification, code generation inside an established codebase, retrieval-augmented answering, most agentic loops — Opus 5 is now the sane default.

What we'd actually do this week

If you're currently running Fable 5 as your default, the practical move is a shadow test rather than a switch: route a slice of real production traffic to Opus 5, keep the outputs, and compare them on your own evals rather than anyone's benchmark. Half a percent on Cursor's hardest setting is a strong signal, but your workload isn't Cursor's benchmark. Two weeks of side-by-side data on your actual tasks will tell you more than any published number — and if it holds, you've halved a recurring cost with a config change.

The broader read: the frontier didn't move this week. The *middle* did, and it moved up sharply enough that the frontier now has to justify itself on price for the first time in a while. We'll keep tracking how the rest of the field responds in [LLM Launches & Updates](https://speka.info/llm-updates/).


What remains unverified

In the interest of not overselling: the Cursor benchmark figure and the computer vision result are both reported results, not independently reproduced ones. Exact per-token pricing, context limits, rate limits on the Pro tier, and availability across enterprise deployment surfaces aren't confirmed here, and we're not going to guess at them. Treat the "roughly half" figure as the directional claim it is — the ratio is the story, and it's worth verifying against your own bill before you rewrite your routing logic.

Frequently Asked Questions

Is Opus 5 better than Fable 5?

No — Fable 5 remains Anthropic's top-intelligence model. Opus 5 performs close to it at roughly half the cost, which makes it the better value for most workloads even though it isn't the more capable model.

How much cheaper is Opus 5 than Fable 5?

Opus 5 costs roughly half of what Fable 5 costs. Exact per-token pricing isn't confirmed here, so treat the 2× ratio as the directional claim and verify against your own usage.

How did Opus 5 score on Cursor's coding benchmark?

On Cursor's coding benchmark at its hardest setting, Opus 5 reportedly scored within half a percent of Fable 5, at about half the cost per task.

Can I use Opus 5 on the $20 Claude Pro plan?

Yes. Opus 5 is available to Claude Pro subscribers at $20 per month, putting capability previously associated with the $200 tier into the entry-level paid plan.

What was the Opus 5 computer vision test?

Anthropic gave Opus 5 an image of a machine part to rebuild as a 3D model but no way to view the image. The model wrote its own computer vision code to extract the shape from raw pixel data and reconstruct the part — a result Anthropic reports no other model completed even after five attempts.

Should I switch my production workload from Fable 5 to Opus 5?

Shadow-test first. Route a slice of real traffic to Opus 5 and compare on your own evaluations — published benchmarks won't match your task mix, and the cost saving only counts if quality holds on your workload.