# Claude Opus 5: Benchmarks, Pricing, Effort Levels and What It Means for Your Business | Alher Tech > Anthropic released Claude Opus 5 on July 24, 2026: number one in the Artificial Analysis Intelligence Index and SWE-bench Verified (97.0%), at $5/$25 per million tokens, half the price of Fable 5. Benchmarks, the five-level effort selector and what it changes for business. - Canonical page: https://alhertech.com/en/blog/claude-opus-5-whats-new/ - Site: Alher Tech (custom software, AI agents and SEO engineering, https://alhertech.com/) - Contact: https://alhertech.com/en/contact/ --- On July 24, 2026, Anthropic released Claude Opus 5, the new generation of its Opus family. It launched day one on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, claude.ai, Claude Code and Claude Cowork, as the default model on Claude Max and the most capable model included in Claude Pro. The headline is economic as much as technical: near-frontier intelligence at half the price of Claude Fable 5. Opus 5 debuts at number one in the Artificial Analysis Intelligence Index and on SWE-bench Verified, while keeping the same $5/$25 per million tokens list price as Opus 4.8. In this guide we break down the benchmarks, the pricing, the new five-level effort selector, where Opus 5 sits in Anthropic's lineup, what changes for developers, and what it means for businesses building with AI. ## What's New in Claude Opus 5 Claude Opus 5 (model ID claude-opus-5) is a text and vision model with a knowledge cutoff of May 2026, the most recent of Anthropic's lineup. The essentials at a glance: - 1M-token context by default: The context window is 1 million tokens out of the box, with up to 128K output tokens (300K on the Batches API with a beta). - A new five-level effort selector: A single parameter with low, medium, high, xhigh and max levels lets you trade cost and latency for capability without switching models. - Adaptive thinking enabled by default: Unlike Opus 4.8, Opus 5 ships with adaptive thinking turned on by default, a breaking change to review before migrating. - Available everywhere from day one: Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, claude.ai, Claude Code and Claude Cowork, with no staggered rollout. ## Benchmarks: How Opus 5 Stacks Up Anthropic's official numbers show a large jump over Opus 4.8 and a lead over both Claude Fable 5 and GPT-5.6 Sol on most fronts: Third-party evaluations back this up. On Vals AI's SWE-bench Verified leaderboard (July 22), Opus 5 scores 97.0%, taking the number one spot. Artificial Analysis measures an Intelligence Index of 61 at max effort, the highest in the world, ahead of Claude Fable 5, with a throughput of 56.3 tokens per second and 64.5 seconds to first token at max effort. One caveat on completeness: LMArena has no ranking for Opus 5 yet, because the model has only been out for a day. Expect that picture to fill in over the coming weeks. - Frontier-Bench v0.1: Claude Opus 5 scores 43.3%, versus 33.7% for Fable 5, 34.4% for GPT-5.6 Sol and 21.1% for Opus 4.8. - GDPval-AA v2: Claude Opus 5 reaches Elo 1861, versus 1747 for Fable 5 and 1736 for GPT-5.6 Sol. - ARC-AGI-3: Claude Opus 5 scores 30.2%, versus 7.8% for GPT-5.6 Sol and 1.5% for Opus 4.8. - OSWorld 2.0: Claude Opus 5 scores 70.6%, beating Fable 5 (66.1%) at a third of the cost. - DeepSWE v1.1: Claude Opus 5 scores 68.8%, behind GPT-5.6 Sol (72.7%) and Fable 5 (69.7%). ## Pricing, Context and Fast Mode Anthropic held the line on price: Opus 5 costs exactly the same as Opus 4.8 and half as much as Claude Fable 5 ($10/$50). - Standard pricing: $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8 and half the price of Claude Fable 5. - Cheaper, easier caching: Cache hits cost $0.50 per million tokens, and the minimum cacheable prompt length drops to 512 tokens. - Fast mode: 2.5x the speed at 2x the price, for latency-sensitive workloads that still need Opus-level quality. - Context and output: 1M-token context window by default and up to 128K output tokens, or 300K on the Batches API with a beta. ## The Effort Selector: Five Levels of Capability The biggest API-level novelty is the new effort selector with five levels: low, medium, high, xhigh and max, with high as the default. More effort means more thinking, more latency and more capability: Artificial Analysis measures an Intelligence Index of 56 at medium effort, 59 at high and 61 at max. There is one restriction to know: disabling thinking is only compatible with effort high or lower, and combining disabled thinking with xhigh or max returns a 400 error. Remember also that adaptive thinking is now enabled by default, a breaking change versus Opus 4.8. For production systems the effort selector is a cost lever: run the long tail of routine requests at low or medium effort, and reserve max for the decisions that justify the extra latency and spend. ## Where Opus 5 Sits in Anthropic's Lineup Opus 5 closes most of the gap to Claude Fable 5's frontier intelligence at half the price. Fable 5 stays ahead on DeepSWE (69.7% vs 68.8%) and remains the reference model for agents on very long-horizon runs. At the top of the range, Claude Mythos 5, available only through Project Glasswing, remains far ahead in offensive cybersecurity and autonomous biology. For everything else, Opus 5 is now the value pick of the frontier tier. Opus 5 also records the lowest misaligned-behavior score of any Claude model to date (2.3), a relevant data point for teams deploying agents in regulated environments. ## What Changes for Developers If you run Opus 4.8 or older Opus models in production, here is what to plan for: - Opus 4.8 moves to legacy: It stays available and serves as the automatic fallback for cyber-related refusals via the server-side-fallback beta. - Claude Opus 4.1 retires on August 5, 2026: Migrate any remaining workloads before that date to avoid disruption. - Not yet available on Opus 5: Priority Tier and server-side web fetch are not supported on Opus 5 for now. - New beta: mid-conversation tool changes: The mid-conversation-tool-changes beta lets you change tools mid-conversation without invalidating the prompt cache. ## What It Means for Businesses Building with AI The practical shift is economic. Coding agents and knowledge-work automation that needed Fable 5 quality now run at Opus prices: near-frontier reasoning at $5/$25 per million tokens changes which automations are worth building at all. The effort selector adds a second lever. Instead of paying top prices for every request, you tune capability per call: routine steps run at low or medium effort, and only the genuinely hard decisions escalate to max, all on the same model and the same integration. As always, published benchmarks are a starting point, not a decision. The numbers that matter are the ones on your own workloads, measured before you commit an architecture to a model. ## How We Put Models Like Opus 5 to Work at Alher Tech We don't build demos: we ship AI agents and automations into production. A release like Claude Opus 5 moves the cost-quality frontier, and we fold it into architectures designed around each client's privacy, cost and reliability constraints. If you are weighing Opus 5 against Fable 5, GPT-5.6 or an open model, we can benchmark them on your real use cases and build the routing and fallback architecture around the winner. - Smart model routing: Opus 5 for the hardest reasoning at half the price of Fable 5, cheaper models for the long tail, and self-hosted LLMs on our own GPUs when privacy or per-request cost is the priority. - Effort-aware cost design: We map each workflow step to the right effort level, so you pay max-effort prices only where the outcome justifies it. - Agents with guardrails: Sandboxed tools, allowlists, spend caps, and human approval on irreversible actions: the scaffolding that makes long-running autonomous agents safe in production. - Evals before production: We benchmark candidate models head-to-head on your real prompts and tasks before committing, so the choice is data-driven, not hype-driven. ## Conclusion Claude Opus 5 lands as the new number one in the Artificial Analysis Intelligence Index and on SWE-bench Verified, at the same price as Opus 4.8 and half the price of Claude Fable 5, with a five-level effort selector that turns capability into a dial rather than a model choice. Fable 5 keeps the edge on DeepSWE and on ultra-long agent runs, and Mythos 5 remains the Project Glasswing specialist, but for most coding and knowledge-work agents Opus 5 is now the default candidate to evaluate first. If you are planning an AI agent or an automation and want to know what Opus 5 changes for your roadmap, we can help you evaluate it and build the production architecture around it. ## Related articles - Claude Opus 4.8: What's New, Benchmarks, Pricing & How to Use It - GPT-5.6 Explained: Sol, Terra and Luna Models, Pricing and Benchmarks - Kimi K3 Explained: The Largest Open-Source AI Model Ever, Benchmarks and Pricing