Claude Opus 5 for coding agents: index 61 at half Fable 5's cost (2026)

Summary. Claude Opus 5, released by Anthropic on 24 July 2026, scores 61 on the Artificial Analysis Intelligence Index, narrowly the highest of any model, one point above Claude Fable 5 (60) and two above GPT-5.6 Sol (59), while costing 26% less per task than Fable 5. It runs at $5 per million input tokens and $25 per million output tokens, the same price as Opus 4.8, and it is now the default model on Claude Max. For teams running coding agents, that combination moves the default choice. This guide covers the benchmarks that matter, the cost-per-task math, and when Opus 5 beats Fable 5, Sonnet 5 and GPT-5.6 Sol.

The short read for a busy engineering lead: if you were paying for Fable 5 to drive everyday coding agents, Opus 5 gives you most of the same quality for a smaller bill, and it is now the model Anthropic points its own Max subscribers at first. The places to keep Fable 5 are narrower than they were a week ago, and this article maps them out.

What changed with Claude Opus 5

Anthropic's own framing is blunt. It calls Opus 5 "a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price." The price held flat from the prior Opus generation: $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8, with a 1 million token context window. Holding the price while lifting performance is the whole pitch, because it widens the set of workloads that are cheap enough to automate.

Three facts define the release. First, Opus 5 is the new default on Claude Max and the strongest model on Claude Pro, so millions of subscribers now hit it without changing a setting. Second, it ships with five effort settings (low, medium, high, xhigh and max), which let you trade tokens for quality per call rather than per model. Third, it is offered in a Fast mode that runs around 2.5 times the default speed at twice the base price, so $10 per million input and $50 per million output when latency matters more than cost.

For a coding agent, the effort dial is the part that changes day-to-day economics. The same model can run cheap on routine edits and expensive on a hard refactor, and Artificial Analysis measured the spread: on GDPval-AA v2, Opus 5's effort levels span 407 Elo points, with output token usage ranging about 8x from low to max effort. You are no longer choosing a model per task so much as choosing an effort level.

The benchmarks that matter for coding agents

Marketing indices are easy to wave away, so start with the coding-specific numbers Anthropic and Artificial Analysis published on launch day.

On Frontier-Bench v0.1, an agentic software-engineering benchmark, Opus 5 surpasses every other model and more than doubles Opus 4.8's score at a lower cost per task. On CursorBench 3.2 at max effort, it lands within 0.5% of Fable 5's peak score while costing half as much per task. On Terminal-Bench v2.1, it hits 89% at max effort, roughly in line with the leader, GPT-5.6 Sol at xhigh effort. And on the Artificial Analysis Coding Agent Index, Opus 5 at xhigh effort with Claude Code is joint first, including the highest score on SWE-Atlas-QnA.

Benchmark	What it tests	Claude Opus 5 result (max/xhigh effort)
Frontier-Bench v0.1	Agentic software engineering	Surpasses all models; more than 2x Opus 4.8 at lower cost per task
CursorBench 3.2	In-IDE coding	Within 0.5% of Fable 5's peak, at half the cost per task
Terminal-Bench v2.1	Agentic terminal use	89%, roughly in line with GPT-5.6 Sol (xhigh)
AA Coding Agent Index	End-to-end coding agents	Joint first (Opus 5 xhigh with Claude Code)
ARC-AGI 3	Novel problem solving	About 3x the next-best model
OSWorld 2.0	Computer use	Beats Fable 5's best result at just over a third of the cost

The pattern is consistent: Opus 5 either matches Fable 5 or gets close, and it does so for less money per task. The one coding-shaped area where it clearly is not the leader is offensive security, which is a deliberate design choice covered below.

The cost math: where Opus 5 actually wins

Token prices alone mislead, because a smarter model can finish a task in fewer turns and tokens. The metric that matters for an agent is cost per completed task, and Artificial Analysis reports it directly for its Intelligence Index workload at max effort.

Model	AA Intelligence Index (max)	Cost per Index task (max)	Notable for
Claude Opus 5	61	$2.03	Highest intelligence; strong intelligence-per-dollar at high effort
Claude Fable 5	60	$2.75 (with fallback)	Flagship; still leads on factual knowledge
GPT-5.6 Sol	59	Not disclosed by AA	Leads Terminal-Bench at xhigh effort
Kimi K3	57	Higher than Opus 4.8	Strong on the AA-Briefcase agentic benchmark
Claude Opus 4.8	56	$1.80	The prior Opus default
Claude Sonnet 5	Below the Opus tier	$1.53	Cheapest per task in this line-up

Read that table as a decision, not a scoreboard. Opus 5 costs $2.03 per Index task against Fable 5's $2.75, the 26% saving Anthropic and Artificial Analysis both lead with. It is still more expensive than Opus 4.8 ($1.80) and Sonnet 5 ($1.53), so it is not automatically the cheapest answer. The real win shows up at high and xhigh effort, where Artificial Analysis found Opus 5 can outperform both Opus 4.8 and Sonnet 5 at a lower cost per task, because it wastes fewer tokens getting to a correct answer. The migration lesson from past model swaps applies here too: the saving is real only if you also right-size the effort setting, or you will pay max-effort prices for medium-effort work. Our note on LLM inference cost for enterprise AI walks through that trap in detail.

Cache pricing is unchanged from recent Opus launches and matters for agents that re-read the same context: cache writes carry a 25% premium at $6.25 per million tokens with a 5-minute time to live, and cache hits get a 90% discount at $0.50 per million tokens. An agent that keeps a large repository or system prompt warm can cut its real input bill well below the $5 headline.

Opus 5 vs Fable 5: when to skip the flagship

The blunt version: for most coding-agent work you can now default to Opus 5 and only reach for Fable 5 on specific jobs.

Skip Fable 5 and use Opus 5 when the work is everyday engineering: in-IDE assistance, pull-request review, bug fixing, multi-file refactors, and terminal or DevOps automation. Opus 5 is within 0.5% of Fable 5 on CursorBench, joint first on the Coding Agent Index, and 26% cheaper per task. Anthropic designed it to be used every day and made it the Max default, which is a strong signal about where it expects the volume to go.

Keep Fable 5 when three things are true. First, factual precision is the product: Artificial Analysis found Opus 5 still trails Fable 5 on the AA-Omniscience knowledge benchmark, and although it gained 7 points over Opus 4.8 on accuracy, it answers more often when uncertain, pushing its hallucination rate up 14 points to 50%. For retrieval-heavy or compliance-sensitive answers, that regression matters. Second, the horizon is very long: Anthropic still positions its flagship for agents that run autonomously for days. Third, the task sits in Fable 5's remaining lead on frontier science, where Artificial Analysis noted Opus 5 matches it on Humanity's Last Exam (53%) but sits behind the GPT-5.6 family on the CritPt physics evaluation.

That is a narrower list than "always use the flagship." For a team spending on Fable 5 purely because it was the best coding model, Opus 5 is the cheaper default that gives up very little.

Opus 5 vs Sonnet 5 and GPT-5.6 Sol

Opus 5 does not exist in a two-horse race. Two other models shape the decision.

Claude Sonnet 5 is the value floor. At $1.53 per Index task it is the cheapest option in the line-up, and for high-volume, latency-sensitive, or budget-capped agents it is often the right call even though it sits below the Opus tier on raw intelligence. The honest framing is a ladder: Sonnet 5 for volume, Opus 5 for everyday quality, Fable 5 for the hardest long-horizon jobs. Our GPT-5.6 vs Claude Sonnet 5 comparison for enterprise agents covers the Sonnet-tier trade-offs.

GPT-5.6 Sol is the closest external rival. It scores 59 on the Artificial Analysis Intelligence Index against Opus 5's 61, and it leads Terminal-Bench at xhigh effort, where Opus 5's 89% is roughly level. The choice between them is rarely about a two-point index gap; it is about which ecosystem your agents, tools and evals already live in, and how each model's effort-versus-cost curve fits your workload. Artificial Analysis noted that at lower effort levels Opus 5 sits just behind the GPT-5.6 family on the intelligence-versus-cost frontier, so if you run agents at low effort for cost reasons, test both rather than assume the higher index wins. For the wider field, our Gemini 3.5 Pro vs GPT-5.6 vs Claude Fable 5 breakdown sets the frontier context.

What early adopters report

Anthropic's launch page carries verbatim notes from early-access customers, which are more concrete than any index.

Sualeh Asif, co-founder of Cursor, said: "Claude Opus 5 delivers near Fable 5 intelligence at Opus speed and cost. On CursorBench it's just under Fable 5 and has many of the same behaviors." That lines up with the published 0.5% CursorBench gap.

Wade Build, CEO of Zapier, reported an end-to-end result: "Claude Opus 5 topped Zapier's AutomationBench leaderboard without spending more tokens than prior Claude models." He added that on a raw account-health workbook it ran a full churn-prevention sequence start to finish where previous models did not pass, and Opus 5 hit 100%.

Ben Kus, CTO of Box, put numbers on enterprise document work: "Box found that Opus 5 outperforms Opus 4.8 by 8% and delivers notable performance gains in the data analysis (11% improvement) and due diligence (17% improvement) workflows." Read together, these are the coding-adjacent, business-automation tasks where the effort-efficient behaviour shows up as fewer failed runs, not just a higher score.

The caveats: hallucination, cyber fallbacks and effort tuning

Three cautions belong in any rollout plan.

The hallucination regression is the important one. Opus 5's 50% hallucination rate on AA-Omniscience, up 14 points from Opus 4.8, means it answers confidently more often when it should abstain. For an agent that writes code, that surfaces as invented APIs or plausible-but-wrong fixes, so keep tests, type checks and retrieval grounding in the loop and do not treat a fluent answer as a verified one. Anthropic's own material stresses that Opus 5 verifies its work and iterates, but design your harness as if it will occasionally be sure and wrong.

Security work is deliberately constrained. Anthropic did not train Opus 5 on cyber tasks and kept it behind its Mythos 5 model on exploit development. Opus 5's cyber classifiers block binary-based vulnerability scanning, penetration testing and exploit generation, and Anthropic expects them to intervene around 85% less often than Fable 5's while still routing flagged requests to fall back to Opus 4.8. If your agents do offensive-security research, plan for those fallbacks or the Cyber Verification Program rather than assuming Opus 5 will comply.

Effort tuning is where teams leave money on the table. Because token usage ranges about 8x from low to max effort, a coding agent pinned at max for every call will erase the 26% saving. Instrument effort per task type and let cheap edits run cheap. On the upside, Anthropic reports Opus 5 as its most aligned model to date, scoring 2.3 on its internal misaligned-behaviour audit, the lowest of its recent models, which lowers the guardrail burden for autonomous runs.

India-specific considerations

For Indian engineering teams and global capability centres, the Opus 5 move is mostly good news, with two local wrinkles.

Cost is billed in US dollars, so the $5/$25 token price and the $2.03-per-task figure carry rupee-exchange exposure that swings your effective spend month to month; teams budgeting in rupees should model that FX drift rather than lock a single conversion. Our guide to Claude's rupee pricing versus US teams breaks down how the billing lands locally. The second wrinkle is data governance: any coding agent that reads customer data or production logs falls under the Digital Personal Data Protection (DPDP) Act 2023, so route agent context through the same consent, minimisation and retention controls you apply elsewhere. Opus 5 keeps the 1 million token context and, consistent with prior Opus models, has no data-retention requirement for general access, which helps, but the obligation to design for DPDP still sits with you.

How to choose: a decision table

Your workload	Recommended default	Why
Everyday coding agents, IDE assist, PR review	Claude Opus 5	Near-Fable-5 quality at 26% lower cost per task
Longest-horizon autonomous agents (days)	Claude Fable 5	Flagship headroom and higher factual accuracy
High-volume, cost-capped, latency-sensitive	Claude Sonnet 5	Lowest cost per task ($1.53) in the line-up
Fact-heavy retrieval where errors are costly	Claude Fable 5	Opus 5 hallucination rate rises to 50%
Offensive-security or exploit research	Mythos 5 or CVP access	Opus 5 blocks exploit generation by design
Terminal and DevOps automation	Opus 5 or GPT-5.6 Sol	89% on Terminal-Bench, near the leader

Treat this as a starting default, then confirm with your own evals on your own repositories, because the only benchmark that pays your bill is the one that reflects your tasks.

FAQ

What is Claude Opus 5?

Claude Opus 5 is Anthropic's Opus-tier model, released on 24 July 2026. It scores 61 on the Artificial Analysis Intelligence Index, the highest of any model, comes close to the flagship Fable 5, and costs $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8.

How much does Claude Opus 5 cost?

Opus 5 costs $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8. A Fast mode runs about 2.5 times faster at twice the price, so $10 and $50. Cache hits are discounted 90% to $0.50 per million tokens, and cache writes carry a 25% premium.

Is Claude Opus 5 better than Fable 5 for coding?

For most coding work it is the better value. Opus 5 lands within 0.5% of Fable 5 on CursorBench 3.2 and is joint first on the Artificial Analysis Coding Agent Index, at 26% lower cost per task. Fable 5 keeps a lead on factual knowledge and the longest autonomous runs.

When should you still use Fable 5 instead of Opus 5?

Keep Fable 5 when factual precision is the product, when agents must run autonomously for days, or for frontier-science tasks. Artificial Analysis found Opus 5 still trails Fable 5 on the AA-Omniscience knowledge benchmark, and its hallucination rate rose 14 points to 50%, which matters for retrieval-heavy or compliance-sensitive answers.

How does Opus 5 compare with GPT-5.6 Sol?

Opus 5 scores 61 on the Artificial Analysis Intelligence Index against GPT-5.6 Sol's 59, and reaches 89% on Terminal-Bench v2.1, roughly level with Sol at xhigh effort. The practical choice usually comes down to ecosystem fit and each model's effort-versus-cost curve rather than the two-point index gap.

What is Opus 5's biggest weakness?

Factual reliability. On the AA-Omniscience benchmark Opus 5 answers more often when uncertain, pushing its hallucination rate up 14 points to 50%, even though accuracy improved 7 points over Opus 4.8. For coding agents, keep tests, type checks and retrieval grounding in the loop so a confident wrong answer gets caught.

Is Opus 5 good for security research?

Only for defensive work. Anthropic did not train Opus 5 on cyber tasks and keeps it behind its Mythos 5 model on exploit development. Its classifiers block binary vulnerability scanning, penetration testing and exploit generation, and flagged requests fall back to Opus 4.8 unless you use the Cyber Verification Program.

Should Indian teams switch to Opus 5?

For everyday coding agents, yes, given the 26% lower cost per task. Two caveats: billing is in US dollars, so rupee-exchange swings affect your real spend, and any agent touching customer data must be designed around the DPDP Act 2023. Model the FX drift and keep consent and retention controls in place.

How eCorpIT can help

eCorpIT builds and runs production AI coding agents and LLM workflows for engineering teams in India and worldwide, including model selection, effort tuning and cost controls of exactly the kind this article describes. We benchmark Opus 5, Fable 5, Sonnet 5 and GPT-5.6 Sol against your own repositories and tasks, then wire the winner into your CI, review and DevOps loops with evals and guardrails in place. If you want to cut agent spend without losing quality, see our work on enterprise AI agent development and LLM migration and cost optimisation, or contact us to scope a pilot.

References

1. Anthropic — Introducing Claude Opus 5
2. Artificial Analysis — Opus 5: Fable 5 level intelligence at a lower cost per task
3. Artificial Analysis — Claude Opus 5 (max): intelligence, performance and price analysis
4. Artificial Analysis — Artificial Analysis Intelligence Index
5. VentureBeat — Anthropic launches Claude Opus 5, a cheaper AI model for coding, agents and enterprise workflows
6. CNBC — Anthropic's Claude Opus 5 AI model rivals Fable 5 and is cheaper
7. Quartz — Anthropic launched Claude Opus 5 at half the price of its most powerful AI model
8. The Decoder — Anthropic's Claude Opus 5 costs well below Fable 5 while matching or beating it across most benchmarks
9. WinBuzzer — Claude Opus 5 targets Fable performance at lower cost
10. Anthropic — Claude Platform pricing

Last updated: 26 July 2026.