Sonnet 5 vs Opus 4.8 vs Fable 5: when to use which

Choosing a Claude model in July 2026 is no longer a single-axis decision. Anthropic ships three frontier tiers at once — Sonnet 5, Opus 4.8, and Fable 5 — and Sonnet 5’s headline $2/$10 introductory price makes it look like the automatic default. It often is. But the honest answer is that the cheapest per-token model is not always the cheapest per-task model, and the gap turns on one lever most teams never touch on purpose: effort.

Sonnet 5 launched on June 30, 2026 as, in Anthropic’s framing, the most agentic Sonnet model yet. Its per-token price is genuinely low — identical to Sonnet 4.6, and roughly half of Opus 4.8. Yet Anthropic’s own tokenizer produces about 30% more tokens for the same text, and independent benchmarking lab Artificial Analysis measured Sonnet 5 at a higher cost per task than Opus 4.8 on their standard-pricing index. Both things are true at once, and the reconciliation is the whole point of this guide.

Below we lay out the three-tier lineup, recompute the pricing math from Anthropic’s published rates, and turn it into two operational tools: an effort-level breakeven matrix and a three-model decision matrix you can act on. Our earlier Sonnet 5 launch coverage framed it as near-Opus quality at roughly half the price; this piece adds the fuller picture — the discount narrows, and can invert, as effort climbs.

For the first time, Anthropic’s frontier is a menu, not a single flagship. Sonnet 5 is the balanced default, Opus 4.8 is the correctness workhorse, and Fable 5 is the general-access, safety- classified version of the Mythos-tier capability. They arrived within five weeks of each other, and their prices span a 5x range from the Sonnet 5 intro rate to Fable 5.

Opus 4.8 shipped on May 28, 2026 with pricing unchanged from Opus 4.7 and a new capability Anthropic calls Dynamic Workflows — Claude plans a large task, fans out potentially hundreds of parallel subagents, and verifies their output against a test suite. Fable 5, from its split-tier launch on June 9, 2026, is the publicly available half of that release.

Start with the published per-million-token rates and work a single, transparent task through each model. The illustrative task below is one million input tokens plus two hundred thousand output tokens, a single pass with no caching, priced at each model’s list rate. The task-cost and ratio columns are our arithmetic, not a vendor figure — recompute them yourself from the row rates.

* Illustrative task = 1M input + 200K output tokens, single pass, no caching. Note the assumption baked in: equal token counts across models. That is exactly where the comparison gets interesting, because real Sonnet 5 workloads do not emit equal token counts.

On raw per-token price, Sonnet 5 is unambiguously cheaper: $6.00 at standard pricing is 60% of Opus 4.8’s $10.00 for the same tokens, and $4.00 during the intro window is just 40%. But Anthropic’s updated tokenizer produces about 30% more tokens for the same text than Sonnet 4.6’s did — a change that also applies to Opus 4.7 and up, Fable 5, and the Mythos models. More tokens per unit of work quietly erodes the per-token advantage.

Artificial Analysis, an independent benchmarking lab, ran the three models through its Intelligence Index at standard (non-promotional) pricing and measured cost per task rather than cost per token. On that basis, Sonnet 5 came in at $2.29 per task versus roughly $1.99 for Opus 4.8 — about 15% higher, despite Sonnet 5’s lower sticker price. The driver was token usage: per the lab, Sonnet 5 used roughly 40% more output tokens per Index task than Sonnet 4.6, and on GDPval-AA it took about three times as many conversational turns. These are Artificial Analysis’s own measurements on their methodology, not a universal cost of running Sonnet 5.

The crucial nuance — and the reason this is not a simple gotcha — is that the $2.29 figure is a blended-effort average at standard pricing. It does not mean Sonnet 5 is always dearer. During the intro window the per-token gap is wider, and at low and medium effort the token expansion is small enough that Sonnet 5 stays genuinely cheaper. The inversion is an effort-dependent phenomenon, not a fixed property of the model.

Here is the part almost no launch coverage operationalizes: the same lab found max effort uses roughly six times the conversational turns of low effort on GDPval-AA. That six-fold swing dwarfs the roughly two-fold difference in per-token price between Sonnet 5 and Opus 4.8. The practical read is that effort level is the primary cost dial and model choice is the secondary one — most teams have it backwards, agonizing over which model to route to while leaving effort at a reflexive default.

The capability ordering is clear and consistent: on the headline agentic-coding benchmark, SWE-bench Pro, Opus 4.8 leads Sonnet 5, which in turn clears Sonnet 4.6 by a comfortable margin. Fable 5 sits highest, though its figure is Anthropic-reported rather than independently audited. The question is whether that capability gap is worth the price gap for your workload — and on many tasks it is not.

Widen the lens beyond a single benchmark and the Sonnet-5-to-Opus-4.8 gap narrows to near-noise on several axes. On agentic knowledge work, they are effectively tied.

The takeaway from the benchmark spread is not that one model wins. It is that Opus 4.8’s advantage is real but concentrated — largest on the hardest coding, negligible on knowledge work where Sonnet 5 sometimes edges ahead. That shape is exactly what makes a per-workload routing strategy pay off rather than a single default.

Sonnet 5 and Opus 4.8 share the same five effort levels — low, medium, high, xhigh, and max — replacing the older manual budget_tokens parameter, which now returns a 400 error on both. That shared dial is what makes a clean crossover analysis possible. The matrix below reads left to right: as effort climbs, the cheaper-in-practice model flips from Sonnet 5 to Opus 4.8. The turn-count column is anchored to Artificial Analysis’s GDPval-AA data; the intermediate cells are qualitative because no source publishes a per-tier dollar figure.

Fold price, benchmarks, and effort together and the guidance collapses into four archetypes. Each maps a workload shape to a model and an effort band. Note that we deliberately leave Fable 5’s effort configuration unspecified — Anthropic has not confirmed it shares the exact five-level dial, so we position Fable 5 by task type rather than by effort tier.

If your work sits at the boundary between coding and pure frontier reasoning, it is worth reading how the top tier compares across vendors — our breakdown of how Fable 5 stacks up against GPT-5.5 covers the cost trade at the frontier, where the per-task math looks different again.

A few traps recur in the coverage of these launches. Two are worth surfacing because they can quietly invalidate a comparison you build in good faith — one about benchmark scales, one about safety behavior that changes what a model will even answer.

The operational conclusion is not to standardize on one model. It is to route by workload and to treat effort as a first-class budget control. A pragmatic default stack looks like this — Sonnet 5 carries the volume, Opus 4.8 handles the hard correctness work, and Fable 5 is held in reserve for genuinely long-horizon tasks.

Two guardrails make this stack hold up. First, batch anything latency-tolerant: the Batch API halves the rate, so Sonnet 5 intro batch runs at $1/$5 per million and Opus 4.8 batch at $2.50/$12.50 — a straight 50% cut on the numbers above. Second, lean on prompt caching for repeated context; a cache read is billed at one-tenth of base input across all three models, which can dominate the economics of any agent that re-reads a large system prompt or codebase.

The larger pattern to watch through the back half of 2026: Anthropic has stopped shipping a single flagship and started shipping a price- segmented ladder, with the tokenizer and effort dial doing more to determine real cost than the sticker price. Teams that build routing and effort control into their agent infrastructure now will adapt to the next tier reshuffle without re-plumbing; teams that hardcode a single model and effort default will keep overpaying every time the ladder shifts. If you want a second pair of hands on that routing logic, our agentic AI transformation engagements start with exactly this kind of per-workload model and effort eval.