What you need to know
- Opus 5 shipped on 24 July 2026 — Anthropic's fourth model release in under two months, after Fable 5, Mythos 5 and Sonnet 5.
- It narrowly leads on intelligence. Artificial Analysis puts Opus 5 (maximum effort) at 61 on its Intelligence Index — effectively tied with Fable 5 (60) and just ahead of GPT-5.6 Sol (59).
- It tops the Agentic Index at 55.3, according to Artificial Analysis — the number that matters most if you are shipping agents rather than chat.
- Pricing is $5/$25 per 1M tokens input/output — the same rate as Opus 4.8, and roughly half Fable 5's $10/$50 per Anthropic's published pricing.
- The context window is 1M tokens, matching Opus 4.8, with up to 128k output tokens.
- Sonnet 5 stays the cheaper Claude Code default — Opus 5 is the reach-for-it option, not the everyday one.
The headline you should care about is not "most intelligent model" — it is "Fable 5-class intelligence at roughly half the token price". Artificial Analysis frames it as comparable intelligence to Fable 5 at about 26% lower cost per task. For a team in Bengaluru or Manchester watching a monthly API bill, that delta is the whole story.
The pricing shift: Fable 5 capability, Opus economics
The reason this launch matters is not a fresh benchmark record. It is that Anthropic has, in effect, folded frontier-grade intelligence back into its standard Opus price band. For most of June, if you wanted the very top of the Claude range you paid Fable 5 rates. Opus 5 collapses that gap.
| Model | Input / 1M | Output / 1M | Context | Where it fits |
|---|---|---|---|---|
| Claude Opus 5 | $5 | $25 | 1M | Hardest reasoning & long-horizon agents |
| Claude Fable 5 | $10 | $50 | 1M | Previous top tier; now premium-priced |
| Claude Sonnet 5 | $3* | $15* | 1M | Default coding-agent workhorse |
*Sonnet 5 standard pricing per Anthropic; a lower launch rate of $2/$10 applied through 31 August 2026. Figures are list prices before prompt caching or Batch API discounts.
Read the table the way a finance-conscious builder reads it. Opus 5 sits at exactly half of Fable 5's per-token cost while, on the independent numbers, matching or edging it on intelligence. Prompt caching still cuts input spend by around 90% on stable context, and the Batch API halves the rate again for non-interactive jobs. So the "true" cost of an Opus 5 run in production is often a fraction of the sticker price — the same optimisation playbook we covered in our guide to cutting LLM API costs applies unchanged here.
What the benchmarks actually say
It is worth being precise, because "tops the Intelligence Index" invites overstatement. According to Artificial Analysis, Opus 5 at maximum reasoning effort scores 61 and is "narrowly the most intelligent model" it has tested — effectively tied with Fable 5 at 60, and ahead of GPT-5.6 Sol at 59 and Kimi K3 at 57. That is a lead measured in a point or two, not a chasm. Treat it as parity at the frontier, delivered at a lower price, rather than a decisive capability jump.
The more meaningful signal is agentic. Artificial Analysis reports Opus 5 topping its Agentic Index at 55.3, ahead of GPT-5.6 Sol and Fable 5, and cites a large lead on its AA-Briefcase agentic-work evaluation. Anthropic's own materials describe state-of-the-art results on coding and knowledge-work suites, while noting Opus 5 still trails the restricted Mythos 5 on cybersecurity tasks. Where a single lab is the only source for a figure, we have attributed it inline rather than presenting it as settled fact — and if you are making a procurement decision, run your own evals on your own workload before trusting any leaderboard.
Why does the agentic number carry more weight than the headline intelligence score? Because the workloads Indian and UK teams are actually monetising in 2026 are agentic ones — coding assistants that open pull requests, support agents that resolve tickets end to end, research agents that plan and execute multi-step tasks. A model that reasons brilliantly in a single turn but loses the thread over a twenty-step tool loop is worth less to those products than one that holds coherence to the end. The Agentic Index is a rough proxy for exactly that endurance, which is why a 55.3 at Opus pricing reads, to a builder, as a bigger deal than a one-point Intelligence Index lead. As always, the caveat holds: a leaderboard measures a fixed task set, and your task set is not theirs.
"We route our planning and multi-step agent loops to whichever model wins on the Agentic Index, not the general one. Opus 5 at 55.3 with Opus pricing means we can finally run the deep-reasoning tier on customer work without the Fable 5 invoice. For a small London studio, that is the difference between offering it and not."
— A Verified Builder · London, UKBuild-vs-buy: what changes for Indian and UK teams
The recurring question for founders across Bengaluru, Pune, London and Cambridge is the same: do we lean on a frontier API, or invest engineering time in a cheaper open-weight stack we host ourselves? Opus 5 nudges that calculation.
On the "buy" side, the maths improved overnight. A halving of the top-tier price narrows the window in which self-hosting an open-weight model pays for itself. If your reason for eyeing a self-hosted GLM or DeepSeek deployment was purely to escape Fable 5 economics, Opus 5 removes some of that pressure — you get comparable intelligence without standing up your own GPU fleet or on-call rota. That said, open weights still win decisively on data residency, per-request unit cost at very high volume, and independence from any single vendor; our write-up on GLM-5.2 topping the open-weight Intelligence Index lays out where that trade still favours self-hosting.
On the "build" side, the interesting shift is agentic coding. With Opus 5's Agentic Index lead available at Opus prices, the deep-reasoning tier becomes viable for the parts of an agent loop that genuinely need it — architectural planning, tricky multi-file refactors, long-horizon task decomposition — while a cheaper model handles the routine turns. That is a routing decision, not an all-or-nothing switch.
Do not read "half the cost of Fable 5" as "cheap". At $25 per 1M output tokens, Opus 5 is still roughly eight times Sonnet 5's output rate. Pointing a chatty agent loop at Opus 5 by reflex is the fastest way to a surprise bill. Default to Sonnet 5 and escalate to Opus 5 only when a task demonstrably needs it.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →Opus 5 or Sonnet 5? A routing decision, not a loyalty test
Anthropic has been deliberate about keeping Sonnet 5 as the default in Claude Code, and that positioning holds after this release. Sonnet 5 was built to bring Opus-like coding, tool use and browser work into a cheaper class, and for the overwhelming majority of coding-agent turns it is the right call. Opus 5 is the model you route to, not the model you live in.
A pragmatic three-tier policy works for most teams:
- Sonnet 5 for the default agent loop — code edits, tool calls, routine reasoning, high-volume turns.
- Opus 5 for the hard 10% — long-horizon planning, gnarly debugging across a large context, and agentic runs where a wrong step is expensive.
- Open weights for anything volume-bound, latency-critical or residency-sensitive, where the per-request economics or compliance posture justify hosting it yourself.
This is the same "split routing" discipline we recommended when Fable 5 launched in June. The models change every few weeks; the routing principle does not.
What builders should do this week
First, re-price your existing Fable 5 workloads. If you moved a pipeline to Fable 5 for the top-tier intelligence, test the same pipeline on Opus 5 — you may hold quality while roughly halving the token bill. Second, identify the one or two places in your agent stack that genuinely need frontier reasoning and wire Opus 5 in there specifically, behind a router, rather than swapping it in globally. Third, keep your evals honest: leaderboard position is a starting hypothesis, not a substitute for measuring accuracy, latency and cost on your own tasks.
For Indian and UK builders in particular, the strategic read is that the frontier is now cheaper to reach than it was in June, which lowers the barrier to shipping genuinely capable agent products from a small team. The differentiator stops being "who can afford the best model" and becomes "who can build the best product around it" — which is exactly the ground independent builders can win on.
One closing note on release cadence. Four flagship models in under two months is not just Anthropic flexing; it is a signal that the model layer is commoditising faster than most roadmaps assume. If your product's moat is "we use the best model", that moat resets every few weeks and rarely in your favour. Build for portability — a clean routing layer, provider-agnostic prompts, and evals you own — so the next release is a config change rather than a rewrite. The teams that treat Opus 5 as one interchangeable option behind a router, not a destination, will absorb the following launch without breaking stride.
Primary sources: Anthropic's launch post at anthropic.com/news/claude-opus-5 and Artificial Analysis's evaluation at artificialanalysis.ai. Verify current pricing on the Claude platform pricing page before you budget.