Anthropic named its new model Opus, which in its own lineup means the middle tier, one rung below the flagship Fable 5. Then it published the benchmarks. On Zapier’s AutomationBench, a test of whether a model can actually drive a chain of real SaaS actions to completion, Claude Opus 5 scored 26.0% against Fable 5’s 17.4%. On FrontierBench v0.1 at maximum effort it hit 43.3% against Fable 5’s 33.7%. On SWE-bench Verified it reached 96.0%. Opus 5 lists at $5 per million input tokens and $25 per million output. Fable 5, the model it beats on those agentic tests, lists at $10 and $50.
So the mid-tier model is cheaper than the flagship and outscores it on the exact workloads most teams are building right now: agents that call tools, operate a computer, and finish multi-step jobs. That is not how model lineups are supposed to work, and it changes how you should route between Anthropic’s models. This is a practitioner’s read on where Opus 5 actually fits, not a launch recap.
What launched on July 24
Opus 5 is Anthropic’s fourth model release in under two months, after Mythos 5, Fable 5, and Sonnet 5. It arrived July 24, 2026, and by launch day Artificial Analysis had ranked it number one on both its Intelligence Index and its Agentic Index, ahead of GPT-5.6 Sol and Grok 4.5.
The hard specs matter for routing decisions, so here they are without the marketing gloss:
- Pricing: $5 per million input tokens, $25 per million output. That is unchanged from the outgoing Opus 4.8, and it is exactly half of Fable 5’s input rate. Batch API work runs at roughly $2.50 and $12.50, the usual 50% discount for non-interactive jobs.
- Context: a 1M-token window, which is both the default and the maximum. Synchronous responses cap at 128K output tokens; the Message Batches API can push output to 300K with a beta header.
- Reasoning: extended thinking is on by default. The
effortparameter dials reasoning depth, and on Opus 5 the meaningful levels sit at the high end (high, xhigh, max). Asking for xhigh or max effort with thinking disabled returns a 400 error, which tells you Anthropic built this model to reason, not to answer fast. - Coding and computer use: SWE-bench Verified 96.0%, SWE-bench Pro 79.2%, SWE-bench Multimodal 59.4%, OSWorld 2.0 (a real desktop-control benchmark) 70.57%.
Anthropic’s own framing is that Opus 5 “approaches the intelligence of Claude Fable 5 at half the price.” Read the benchmark table and that is understated. On agentic tasks it does not approach Fable 5; it passes it.
The naming no longer maps to capability
For a year, Anthropic’s tiers were legible, and if you are new to the family the Claude platform guide lays out how they used to stack. Sonnet was the cheap workhorse, Opus was the capable middle, and the flagship (Fable 5, released June 10) was the ceiling you paid a premium to reach. Buyers could route on the name.
Opus 5 breaks that. Line up the three current models on the work most teams are actually shipping:
| Model | Input / Output ($/M) | SWE-bench Verified | SWE-bench Pro | AutomationBench | FrontierBench (max) |
|---|---|---|---|---|---|
| Sonnet 5 | $2 / $10 (intro), then $3 / $15 | high-70s% | mid-70s% | lower | lower |
| Opus 5 | $5 / $25 | 96.0% | 79.2% | 26.0% | 43.3% |
| Fable 5 | $10 / $50 | high-90s% | 80.0% | 17.4% | 33.7% |
The one place Fable 5 still leads is SWE-bench Pro, the harder, longer-horizon coding benchmark, and it leads by eight tenths of a point (80.0 versus 79.2). Everywhere else in the agentic column, the half-price model wins. If your workload is agents driving tools and computers rather than solving the single hardest coding problem in one shot, Fable 5 is now the wrong default. You would be paying double for a model that scores lower on your task.
That is the practitioner headline. The tier name tells you the price. It no longer reliably tells you the capability.
Where each model actually fits
Pricing sets the shape of a sane routing policy. Here is how I would wire it for a team paying its own token bill.
Sonnet 5 is the default and the high-volume floor. At $2/$10 on introductory pricing through August 31, then $3/$15, it is less than a third of Fable 5’s rate and it holds up on well-scoped work: classification, extraction, routine code edits, retrieval-augmented answers, the first pass on almost anything. Most of your call volume should land here. Sonnet 5 also carries the full effort dial (low through max), so you can push it harder on the tasks that are close calls before you escalate.
Opus 5 is the new escalation target and, for agent work, arguably the new default. When a task involves a tool-calling loop, a computer-use step, or a multi-step job where a wrong turn compounds, Opus 5 is where the agentic benchmark lead lives, at half of Fable 5’s price. The old pattern was “escalate hard cases to the flagship.” The new pattern is “escalate hard cases to Opus 5, and only reach past it when you have a specific reason.” For a lot of production agent stacks, Opus 5 quietly becomes the model you run all day.
Fable 5 is now a narrow specialist, not the ceiling. Reserve it for the band where it still leads: the hardest long-horizon coding, the SWE-bench Pro class of problem, the one-shot solve on a genuinely difficult task where the extra points are worth double the token cost. That is a real band, but it is smaller than the flagship’s price implies. Sending your whole agent workload to Fable 5 out of habit is now a line item you can cut without losing capability on most of it.
A concrete escalation ladder: Sonnet 5 handles the request; if it fails a self-check, times out, or hits a confidence floor, retry on Opus 5; if Opus 5 stalls on the specific hard-coding class where Fable 5 leads, escalate again. Three tiers, each earning its place, with the expensive one guarded.
The effort dial is the second axis, and it multiplies the bill
Model choice is only half the cost story. Every current Claude model ships a per-request reasoning dial, and on Opus 5 that dial runs high, xhigh, and max, with thinking on by default. Reasoning tokens are billed as output. Push a batch of agent calls to max effort and you can multiply the output-token cost of the same job several times over without changing the model name on the invoice.
The discipline that separates teams with a predictable AI bill from teams with a surprise one is treating effort as a routing decision, not a global default. Route the bulk of Opus 5 traffic at its floor effort and lift only the calls that demonstrably need it, the same way you route the bulk of volume to Sonnet 5 and lift only what fails. Anthropic building Opus 5 to reason by default is a capability win and a cost trap in the same feature. The teams that instrument per-task effort will get the capability; the teams that leave it on max will get the trap.
The caveats worth keeping
Two of them. First, launch-day benchmarks are a snapshot from the vendor’s preferred test set, and single-effort-level numbers can flatter a reasoning-heavy model that burns tokens to get them. FrontierBench 43.3% is a max-effort figure; the price of that score is reasoning tokens you pay for. Judge Opus 5 on cost per completed task in your own workload, not on the leaderboard row.
Second, SWE-bench Pro is not a rounding error. If your core use case is the hardest class of autonomous coding over long horizons, Fable 5’s narrow lead there is the reason it still exists in the lineup, and the broader agentic coding-tool field is worth weighing before you commit a stack to any one model. Test both on your actual repository before you decide the half-price model wins your specific job. The point is not that Opus 5 beats Fable 5 everywhere. It is that the price tag stopped predicting which one wins, and you now have to measure.
For a telecom running enterprise agents, the operational takeaway is blunt: Anthropic just cut the price of frontier-class agentic capability in half and renamed it a middle tier. The buyers who route on the name will overpay for Fable 5 on work Opus 5 does better. The buyers who route on measured cost per task will quietly move most of their agent traffic down a tier and keep the capability. That gap, between routing on the label and routing on the number, is where the token savings for the back half of 2026 will come from.
Sources: Anthropic Claude Platform Docs, “What’s new in Claude Opus 5”; MarkTechPost, “Meet the New Claude Opus 5” (July 24, 2026); Artificial Analysis Intelligence Index and Agentic Index; Codersera, “Claude Opus 5 vs Sonnet 5: Which to Use in 2026”; Finout, “Claude Sonnet 5 Pricing 2026”.