Claude Opus 5 Is Here: Frontier-Level Reasoning at Half the Price of Fable 5

On July 24, 2026, Anthropic released Claude Opus 5, its fourth model launch in two months and the clearest signal yet of how fast the pace of frontier AI releases has become. The headline claim — intelligence that "comes close to the frontier" of Anthropic's own top-tier Fable 5 model, at exactly half the price — is the kind of statement that would normally invite skepticism. The benchmarks Anthropic published alongside the release, including a perfect score on the 2026 International Mathematical Olympiad and a new state-of-the-art result on ARC-AGI-3, make it harder to dismiss.

For businesses that have spent the past two months tracking Anthropic's unusually fast release cadence — we covered Claude Sonnet 5's launch at the start of July and Claude Cowork's mobile expansion two weeks later — Opus 5 is the release that resets the top of the model lineup. Here is what actually changed, and what it means for choosing between Claude models in production.

What Anthropic Actually Shipped

Opus 5 is priced identically to its predecessor, Opus 4.8: $5 per million input tokens and $25 per million output tokens. A Fast mode, running at roughly 2.5 times default speed, is available at twice that base price. The context window grows to 1 million input tokens with 128,000 tokens of output by default, extendable to 300,000 output tokens on the Batch API through a beta header. Opus 5 is now the default model on Claude Max and the strongest option on Claude Pro, and it is available to developers immediately as claude-opus-5 on the Claude API and in Claude Code.

The pricing detail that gives the release its headline is the comparison to Fable 5, Anthropic's most capable and most expensive model, priced at $10 per million input tokens and $50 per million output tokens. Opus 5 costs exactly half that on both counts, while — according to Anthropic's own benchmark disclosures — closing most of the capability gap.

The Benchmarks Are Unusually Hard to Wave Away

Model launches routinely lead with cherry-picked benchmark wins. Two results from this one are worth taking seriously on their own terms.

A perfect score on the 2026 International Mathematical Olympiad. Anthropic ran Opus 5 against all six 2026 IMO problems with no tools and no agent harness — just the model reasoning on its own. A three-model judge panel scored all 24 generated solutions correct, and human experts independently graded one pre-specified solution per problem at 7 out of 7, for a final score of 42 out of 42. The gold-medal cutoff for the competition was 29 out of 42. This is not a benchmark Anthropic built internally; it is a fixed, externally set human competition, which makes the result meaningfully harder to dispute than most AI capability claims.

A new state-of-the-art result on ARC-AGI-3. As of July 24, 2026, Opus 5 (High) leads the ARC-AGI-3 leaderboard at 30.2%, more than twenty times Opus 4.8's 1.5% and ahead of every other model tested, including five public demo environments no model had previously solved. ARC-AGI is specifically designed to resist memorization and brute-force pattern matching, which is why labs and independent researchers treat it as one of the more credible signals of genuine reasoning ability rather than benchmark saturation.

On more applied measures, the pattern holds. Opus 5 scores 43.3% on Frontier Bench v0.1 against Fable 5's 33.7% and more than double Opus 4.8's 18.7%. On CursorBench 3.2, a coding-agent benchmark, it lands within half a percentage point of Fable 5's peak score — at half the cost per task. Anthropic also reports Opus 5 as its most aligned model to date, scoring lowest among recent Claude models on its automated behavioral audit for deceptive or misaligned behavior, with no new concerning capabilities identified in the accompanying system card.

Why the Pricing Move Matters More Than Any Single Benchmark

Anthropic could have priced Opus 5 as a premium tier above Opus 4.8 and still had a defensible release, given the benchmark jump. Instead it held pricing flat and let the capability gain fall straight to the bottom line for anyone already budgeting for Opus-class work. That is a deliberate positioning choice, not an accident.

It also sharpens a decision every business running Claude in production now has to make explicitly, across three tiers rather than two:

• Fable 5 ($10/$50) — the ceiling on raw capability, for the narrow slice of tasks where the last few points of reasoning quality justify the cost.
• Opus 5 ($5/$25) — frontier-class reasoning and agentic coding at half that price, now the rational default for most demanding production workloads that previously required Fable-tier spend.
• Sonnet 5 ($2–3/$10–15, per our earlier coverage) — the volume tier for high-throughput, sustained tool-use work like support agents and document processing.

The practical effect for most SMEs is that the "we need Opus-level quality but can't afford Opus-level cost" trade-off, which we have heard from clients evaluating agent deployments all year, has just gotten meaningfully easier to resolve. Work that was shelved at Fable 5 pricing, or run on Sonnet 5 as a cost compromise despite needing more reasoning depth, is worth re-testing against Opus 5 before assuming the economics still don't work.

What This Means If You're Building on Claude

Re-benchmark before you re-platform. A 20x jump on ARC-AGI-3 and a perfect IMO score are compelling headline numbers, but they describe frontier reasoning and mathematics, not necessarily your specific document-extraction pipeline or customer-support workflow. Run your own evaluation set against Opus 5 before migrating a production workload off Sonnet 5 or Opus 4.8 — the tasks that benefit most from the reasoning jump are the ones with genuinely hard, multi-step logic, not high-volume, low-ambiguity work that Sonnet 5 already handles well.

Treat the Fable 5 vs. Opus 5 choice as a per-workload decision, not an account-wide default. With Opus 5 covering most of Fable 5's capability at half its price, the remaining cases that justify Fable 5 spend are narrower than they were a week ago. Segment your workloads by how much the marginal reasoning quality is actually worth, rather than defaulting every task to your most expensive available model.

Watch the cadence, not just the model. Four releases in two months — Sonnet 5, Cowork's expansion, and now Opus 5 — is a pace that rewards businesses with a lightweight, repeatable process for evaluating new models against their own workloads, and penalizes those that re-architect around a specific model version. The context engineering discipline we've written about before is exactly the kind of investment that pays off regardless of which model sits behind it next month.

The Bottom Line

Claude Opus 5 is a genuine step change dressed up as a routine point release: a perfect score on one of the hardest fixed human benchmarks in existence, a new state-of-the-art result on a benchmark specifically designed to resist gaming, and all of it priced at exactly half of Anthropic's own flagship model. For Swiss SMEs, the immediate action isn't to switch models reflexively — it's to re-run the cost-versus-capability math on any workload that was previously judged "not worth Opus-tier spend" and see whether that judgment still holds.

If you're weighing which Claude model tier actually fits your workloads — or want a second opinion before committing engineering time to a migration — get in touch with our team. The right model choice this month is rarely the same one that was right two releases ago.