0 and $50. There is a Fast mode at roughly 2.5x the speed for 2x the base price if latency is your constraint.
On benchmarks it is not a small gap. Opus 5 posts 43.3% on Frontier-Bench against Fable 5's 33.7%, edges ahead on GDPval-AA v2 (1861 vs 1747), and lands within a point on SWE-bench Pro. Fable 5 still wins a few — DeepSWE (69.7 vs 68.8) and the held-out Legal Agent benchmark (13.3 vs 11.7) — but the pattern is clear: the cheaper model matches or beats the flagship across most of the board.
What I read into that is not "Fable 5 is obsolete." It is that the price-to-capability curve just bent, and if you are still defaulting to the most expensive model out of habit, you are leaving money and speed on the table.
The Three Things That Actually Matter for Real Work
Benchmarks are a starting point, not a decision. On real engagements, three properties decide which model I reach for.
Instruction-following and self-verification. This is where Opus 5 genuinely pulled ahead. Anthropic built it for bounded professional tasks, and it is noticeably better at doing exactly what you asked and then checking its own work before it stops. That matters more than raw intelligence for the kind of work I do — a validation-rule rewrite or a data reconciliation is only useful if the model actually holds to the spec and verifies the result. A legal team reported 26% fewer tokens than the previous Opus at max reasoning for the same quality; a financial-modeling team measured 9 points more accuracy with a third fewer tool calls. Discipline, not just horsepower.
Design and creative taste. Fable 5 still wins here, and it is not close. When the deliverable is a client-facing portal, a branded Experience Cloud page, or a pitch deck, Fable's judgment about layout, restraint, and what looks intentional is worth the premium. Opus 5 is competent; Fable 5 has taste.
Cost per job, not cost per token. This is the trap. Opus 5 is half the price per token, so it looks like an automatic win. But on open-ended, ambiguous tasks it will happily iterate — spin up more checks, run more passes — and a model that is half the price per token but burns three times the tokens is not cheaper. Cheaper per token is not the same as cheaper per job. The savings are real on bounded work with a clear finish line. On loose exploration, measure before you assume.
When to Reach for Which
Here is how I am routing work right now:
| Task | Reach for | Why |
| --- | --- | --- |
| Flow and validation-rule analysis, bounded and checkable | Opus 5 | Strong instruction-following and self-verification; you can define pass/fail, and it is half the price |
| Data-integrity checks, migration reconciliation | Opus 5 | Bounded, verifiable, and token-efficient at low to medium effort |
| Client-facing UI, decks, brand-sensitive design | Fable 5 | Still the better taste and creative judgment |
| Open-ended discovery, ambiguous research | Fable 5, or measure both | Opus can over-iterate and burn tokens when the goal is loose |
| Routine drafting, summaries, first-pass copy | A smaller Sonnet-class model | Opus and Fable are both overkill here; match intelligence to the task |
The row that surprises people is the last one. A lot of daily knowledge work — meeting summaries, first-draft copy, straightforward extraction — does not need a frontier model at all. A cheaper Sonnet-class model does it well, and reserving the expensive models for the jobs that need them is most of good cost control.
What Hasn't Changed
Fable 5 is still the stronger model when you need its ceiling — the ambiguous, long-horizon, taste-heavy work is where it earns its price. That has not moved.
Neither has the uncomfortable truth underneath all of this: these models are non-deterministic. Same prompt, same model, two different runs, and you can get meaningfully different behavior — one run follows your instruction to the letter, another wanders. No benchmark captures that. Which is why the thing that moves outcomes most is not the model badge at all. It is context and instruction: how clearly you define the goal, what you feed in, and whether you gave the model a way to check that it is actually done. I have watched a well-prompted cheaper model beat a badly-prompted expensive one more times than I can count.
What to Do Right Now
1. Benchmark on your own workflows, not the leaderboard. Run your real tasks — your Flow analysis, your migration checks — through both models and compare. The public numbers point you at a hypothesis; your work confirms it.
2. Define pass/fail so the model can verify itself. Give it a concrete stopping condition: a SOQL check that must return zero rows, a test that must pass, an explicit list of criteria that must all be green. This is exactly where Opus 5's verification edge converts into fewer bad handoffs.
3. Use a cheap orchestrator that only delegates. Keep the driver model light and let it hand bounded work to the right model per step. It protects both your budget and your session limits, and it keeps the expensive intelligence pointed at the parts that need it.
4. Match the intelligence to the task. Reserve the frontier model for the jobs that actually need a frontier model. Most knowledge work does not, and pretending it does is just a more expensive way to get the same result.
Not sure which model belongs on your Salesforce automation work? Let's talk — I will help you match the model to the job, not the hype.