Claude Opus 5 vs Fable 5: Is Half the Price Worth It?

Anthropic launched Claude Opus 5 on July 24, 2026, and the question every team asked within an hour was the same one: if Opus 5 gets close to Fable 5 at half the price, why would anyone keep paying for Fable 5?

The numbers are simple. Opus 5 costs $5 per million input tokens and $25 per million output tokens. Fable 5 costs $10 and $50: an exact 2x gap on both sides, with no small print about tiers or minimums. Anthropic’s own launch claims say Opus 5 lands within 0.5% of Fable 5’s peak on CursorBench 3.2 and surpasses Fable 5 on OSWorld 2.0 at one-third the cost.

If you want to run the comparison yourself, you can send the same prompt to both model IDs from Apidog and diff the responses and token counts side by side.

The two models at a glance

Two rows deserve a second look before anything else.

The cutoff row cuts against the usual “newer tier is cheaper, older tier is smarter” story. Opus 5 knows about the world through May 2026. Fable 5 stops at January 2026, the same cutoff as Sonnet 5. Four months of package releases, framework versions, and API changes sit inside Opus 5’s training data and outside Fable 5’s. If your work touches fast-moving libraries, the cheaper model is also the fresher one.

The positioning row cuts the other way. Anthropic still describes Fable 5 as its most capable widely released model, and it has not moved that label onto Opus 5. Opus 5 is the recommended default, which is a different claim: it is what you should reach for when you are unsure, not what you should reach for when the task is at the edge of what any model can do.

For the full spec sheet on the newer model, see our Claude Opus 5 explainer. For Fable 5’s launch details, see what Claude Fable 5 is.

What Anthropic actually claims

Here are the launch claims that bear on this comparison, stated as Anthropic states them.

Now the caveat, because it is the whole ballgame. Every one of these figures is vendor-run. Anthropic designed the evaluation setup, chose the effort settings, and published the results. As of July 25, 2026, none has been independently reproduced. That is not an accusation of bad faith; it is the normal state of a launch on day two. But “within 0.5% of Fable 5” describes a specific harness at a specific configuration, and the gap you measure on your own tasks may be larger or smaller.

The OSWorld claim is the more interesting of the two head-to-head numbers, because it is the one where Opus 5 is said to beat Fable 5 rather than approach it. Computer-use tasks reward persistence and recovery over raw reasoning depth, which fits a model tuned for agentic loops. The CursorBench claim is the softer one: “within 0.5% of Fable 5’s peak” means Fable 5 still holds the peak.

We break every number down, including what each benchmark does and does not measure, in the Opus 5 benchmarks deep dive. Fable 5’s own launch figures are in the Fable 5 benchmarks post.

The case for Opus 5

The price gap is flat and it compounds. At $5/$25 against $10/$50, a workload costing $3,000 a month on Fable 5 costs $1,500 on Opus 5 for identical token volume. No threshold or commitment narrows it.

The cost levers stack on top of the halving. Opus 5’s Batch API rate is $2.50/$12.50, a further 50% off. Cache hits and refreshes run $0.50 per million, and the prompt caching minimum dropped to 512 tokens (from 1,024 on Opus 4.8), so short system prompts that were previously too small to cache now cache with no code change. Tool-use system prompt overhead is lower too: 286 tokens with tool_choice set to auto or none, against 290 on Opus 4.8 and 675 on Opus 4.7. Together these move a real bill. Our Opus 5 pricing breakdown has the full table, and cutting your Claude API bill covers the general levers.

You can spend up when a task deserves it. Opus 5’s output_config.effort parameter defaults to high and the levels were recalibrated at launch, with low and medium meaningfully stronger than on previous Opus models. That gives you a dial rather than a fixed rate: cheap classification at low, hard agentic coding at xhigh. The effort parameter guide explains the sweep to run instead of carrying old settings over.

It is the fresher model, by four months.

It is where the defaults are going. Opus 5 is the new default for Max subscribers and the highest-performing model available to Pro users, and it ships on the Claude API, claude.ai, Claude Code, Claude Cowork, Amazon Bedrock, Claude Platform on AWS, Google Cloud, Microsoft Foundry, and GitHub Copilot. Anthropic’s models overview says plainly that if you are unsure which model to use, start with Opus 5.

The case for Fable 5

Steelmanning the expensive model is harder work, so let’s be specific rather than polite.

The capability designation has not moved. Anthropic could have crowned Opus 5 its most capable widely released model on launch day. It did not. When a vendor keeps a superlative on the older, pricier model while shipping a cheaper one it is loudly proud of, that is a signal about the top of the capability curve, not a marketing oversight. On the hardest problems, the ones where a single failed run costs more than a month of API spend, Fable 5 is still the model Anthropic points at. Its launch announcement is the reference for that positioning.

Always-on thinking removes a failure mode. Fable 5 thinks on every request: Anthropic’s model comparison lists its adaptive thinking as “always on,” and it is the only current model described that way. Opus 5 lets you disable thinking, and Anthropic’s own prompting guidance documents what happens when you do: tool calls occasionally get written out as plain text instead of being executed (and that leaked text then poisons later turns in an agentic loop), and internal XML tags sometimes appear in visible output. Anthropic’s recommended mitigation is to keep thinking on and control cost with lower effort instead. A model that never offers you the footgun cannot be shot by it.

Opus 5’s flexibility has sharp edges. Combining thinking: {type: "disabled"} with effort xhigh or max returns a 400, enforced per request. Thinking now runs by default when you omit the thinking field, so max_tokens budgets sized for a non-thinking workload can truncate. Opus 5 also delegates to subagents more readily and produces longer default responses, both of which quietly raise cost if you do not constrain them. Fable 5 carries none of this migration surface. If you do move, our Opus 4.8 to Opus 5 migration guide has the breaking changes in full.

Switching costs are not zero. Your evals, prompts, effort settings, and cost model are calibrated against the model you run today. “Half the price” is a per-token statement, not a project-cost statement. If migration takes two engineers a week and the workload is small, the arithmetic does not work. The Fable 5 vs Opus 4.8 comparison covers what the Fable tier buys over the Opus tier generally, and Fable 5 pricing covers its cost structure.

What the price gap looks like in real requests

Multipliers are easy to nod at and hard to feel. Here is one agentic coding task: 200,000 input tokens (a large repository context plus a long tool-call history) and 20,000 output tokens.

At 500 tasks a month, the gap is $750. That is inside the noise of most engineering budgets, and if Fable 5 saves your team one failed run a month, it pays for itself. At 5,000 tasks a month, the gap is $90,000 a year, which is a headcount conversation.

So the decision rule is about volume, not about which model is smarter. Below a few hundred hard tasks a month, pick on capability and ignore the price. Above a few thousand, you need evidence Fable 5 is better on your workload, not on Anthropic’s harness.

One more piece of arithmetic: Opus 5’s fast mode (2.5x output speed, a research preview on the first-party API only) is priced at $10/$50, exactly Fable 5’s base rate. If latency is the constraint rather than capability, you can spend Fable-5 money on a faster Opus 5 instead of a slower Fable 5. Note that inference_geo: "us" adds a 1.1x multiplier across every token category on Opus 5, shifting these numbers by 10%.

A decision framework by workload

The pattern: Opus 5 wins on throughput economics, Fable 5 wins on tail risk. If you run many tasks and can tolerate an occasional weaker result, halving your bill is the obvious move. If you run few tasks and a single bad result is expensive, the price difference is not what you should be optimizing.

Running the comparison on your own workload

Vendor benchmarks tell you what happened on someone else’s tasks. The number that decides this for you comes from your own evaluation set, and getting it takes an afternoon. Both models sit behind the same Messages endpoint, so the comparison is a one-string change:

curl https://api.anthropic.com/v1/messages \
  --header "x-api-key: $ANTHROPIC_API_KEY" \
  --header "anthropic-version: 2023-06-01" \
  --header "content-type: application/json" \
  --data '{
    "model": "claude-opus-5",
    "max_tokens": 8192,
    "output_config": {"effort": "high"},
    "messages": [
      {"role": "user", "content": "Refactor this handler and explain the tradeoffs."}
    ]
  }'

Swap claude-opus-5 for claude-fable-5 and you have the other side of the test. A practical setup in Apidog looks like this:

1. Store the API key as an environment variable so it never sits in a request body, and create one environment per model with the model ID as a variable.
2. Save the request in a collection and run the same scenario against both environments, so the only thing that changes is the model string.
3. Read the usage fields on each response. Compare input, output, and cache-read tokens; that is where the real cost difference shows up, since Opus 5 tends toward longer default responses.
4. Inspect the SSE stream and tool-call payloads when you compare agentic behavior. Whether a model executed a tool or wrote the call out as text is visible in the raw event stream.
5. Repeat the Opus 5 run at low, medium, high, and xhigh effort. The real question is usually not “Opus 5 or Fable 5” but “which effort level on Opus 5 matches the Fable 5 output I am paying for.”

Step five is the one most teams skip and the one that decides the money. Download Apidog to set the environments up, or follow the full request walkthrough in our Opus 5 API guide. Anthropic’s pricing page has the authoritative rates for both models.

The honest ceiling

Neither model in this comparison is the top of Anthropic’s stack.

Fable 5 remains the most capable widely released model by Anthropic’s own designation, and Opus 5 has not taken that label. And above both of them, on two specific capability axes, sits Mythos 5: Anthropic states that Opus 5 still trails Mythos 5 on cybersecurity exploitation and autonomous biology research. Mythos 5 is the classifier-free sibling released to Fable 5 alongside a limited program rather than to the general public, so it is not an option for most teams; our Fable 5 vs Mythos 5 comparison covers eligibility.

The accurate summary of Opus 5 is therefore narrower than the headlines: frontier-class capability at half the frontier price, with a named ceiling above it. That is a strong product. It is not “the best model now costs half as much,” and anyone selling you that version is skipping the part where Anthropic itself kept the capability crown on the pricier model.

FAQ

Is Claude Opus 5 better than Fable 5? On Anthropic’s own benchmarks, Opus 5 surpasses Fable 5 on OSWorld 2.0 computer-use tasks and lands within 0.5% of Fable 5’s peak on CursorBench 3.2, both at a fraction of the cost. Those figures are vendor-run and unreproduced as of July 25, 2026. Anthropic still designates Fable 5 as its most capable widely released model, so on the hardest tasks Fable 5 retains the edge.

How much cheaper is Opus 5 than Fable 5? Exactly half on both sides: $5/$25 per million input/output tokens versus $10/$50. Opus 5’s Batch API rate of $2.50/$12.50 widens the gap further for non-urgent work.

Which model has the newer knowledge cutoff? Opus 5, by four months. Its reliable cutoff is May 2026; Fable 5’s is January 2026.

Should I migrate my Fable 5 workload to Opus 5? Run the numbers on volume first. Below a few hundred hard tasks a month the savings rarely justify the migration work. Above a few thousand they clearly do, provided your own evaluation set confirms the quality holds.

What is the model ID for each? claude-opus-5 and claude-fable-5, both with no date suffix. On Amazon Bedrock, Opus 5 is anthropic.claude-opus-5.

Is there anything Opus 5 supports that Fable 5 does not? Opus 5 exposes the output_config.effort parameter with recalibrated levels, accepts mid-conversation system messages, supports fast mode at $10/$50, and caches prompts as short as 512 tokens. One gap in the other direction: Priority Tier is not supported on Opus 5, though Opus 4.8 keeps it.

The short answer

Half the price is worth it for most teams, and the reason is volume rather than capability. Anthropic’s numbers put Opus 5 close enough on agentic coding and ahead on computer use that the 2x saving compounds into real budget at thousands of tasks a month. Add the fresher May 2026 cutoff and the effort dial and the default answer is Opus 5.

Keep Fable 5 when the task sits at the edge of what any model can do, when a single failed run costs more than the price difference, or when your volume is low enough that migration work outweighs the savings. And treat every benchmark here as what it is: a vendor’s own measurement, on a vendor’s own harness, waiting for someone to reproduce it. Your evaluation set is the tiebreaker.