Claude Opus 5 is Here: Near-Fable 5 Intelligence at Half the Cost
More Than a Point Release
On July 24, Anthropic launched Claude Opus 5 — the latest in the Opus line, the workhorse model most developers default to when using the Anthropic API. Within five hours, the Hacker News thread hit 1,106 points and 590 comments.
The real story is the pricing strategy: Opus 5 delivers near-frontier model Fable 5 performance — which costs $10/$50 per million tokens — at $5/$25. Half the price for roughly the same intelligence.
Same cost as Opus 4.8, double the Frontier-Bench performance. Anthropic is shipping "more for the same money" instead of raising prices.
Benchmarks: Leading Across the Board
Per BenchLM, Opus 5 scores 85.88/100 overall, ranking #1 out of 215 tracked models:
| Model | BenchLM Score | Price (input/output per 1M tokens) |
|---|---|---|
| Claude Opus 5 | 85.88 (#1) | $5 / $25 |
| Claude Mythos 5 | 83.01 (#2) | N/A (Trusted Access) |
| Claude Fable 5 | 82.76 (#3) | $10 / $50 |
| GPT-5.6 Sol | 81.46 (#4) | $5 / $25 |
| Kimi K3 | 79.98 (#5) | free |
| Claude Opus 4.8 | 77.44 (#6) | $5 / $25 |
Three standout results:
SWE-bench Verified: 97% (#1). The benchmark closest to real developer work — fixing bugs in actual Python codebases like Django, Flask, and scikit-learn. Opus 5 leads the board, ahead of GPT-5.6 Sol (96.2%) and Fable 5 (95%).
Frontier-Bench v0.1: 43.3% — double Opus 4.8. The hardest agentic coding benchmark, requiring models to plan, write code, debug, and iterate. Opus 5 beats Opus 4.8 at a lower cost per task.
CursorBench 3.2: 0.5% behind Fable 5 at half the cost. For Cursor and Claude Code users, the signal is clear: near-Fable quality without the premium.
Beyond coding, Opus 5 leads Knowledge (#1 out of 53 models, score 93.5), ranks third in Agentic (#3/129, 69.4) and Multimodal (#3/32, 89.1). The one soft spot: pure Coding ranks #9/130 — Anthropic is reserving Fable 5 for the absolute peak.
The Real Differentiator: Self-Verification
What sets Opus 5 apart from Opus 4.8 isn't just benchmark scores. It's how the model behaves.
On Frontier-Bench, Opus 5 was handed a technical drawing and asked to write code that reconstructed a 3D CAD model. The catch: the model had no direct way to view the drawing. Its response: write its own computer vision pipeline to extract geometry from raw pixels, then reconstruct the entire machine part. No other model solved this after five attempts — Opus 5 did it repeatedly.
An engineer at a trading firm used Opus 5 to build a market data feed for a new exchange in a single session. Previous models couldn't complete this at all, even with detailed plans. Opus 5 also built its own test harness to validate the code — since no live feed was available.
An early-access user described: "During a rearchitecting session, Opus 5 pushed back on a design I proposed, and it didn't fold when I insisted. Instead, it explained exactly what was valuable in my idea, narrowed its objection to a single design question, and proposed a compromise that kept the good part while fixing the flaw."
This isn't the familiar "AI says yes" pattern. Opus 5 has opinions — exactly what Anthropic means by "thoughtful and proactive."
Token Efficiency: Doing More With Less
A critical detail for developers watching costs: Opus 5 consumes noticeably fewer tokens than Opus 4.8. One legal-sector customer reported 26% fewer tokens at the same reasoning level. A financial firm noted "a third fewer tool calls and 60% less processing time."
Combined with unchanged pricing, real cost-per-task drops — not just on paper.
Anthropic also introduced effort control: choose low/medium/high/xhigh/max effort depending on the task. At low effort, Opus 5 passes more tasks than any other model on Zapier AutomationBench. At max effort, it hits peak scores without burning tokens at Fable 5 rates.
What This Means for Developers
A free upgrade — no action needed
Opus 5 is the default model on Claude Max and the strongest on Claude Pro. Anyone on Opus 4.8 via API or Claude Pro keeps the same budget. More quality, same price.
Fable 5 is now for edge cases only
With Opus 5 hitting 97% on SWE-bench, most daily coding tasks don't need Fable 5. Reserve it for tasks demanding absolute precision or the most complex reasoning. That's a 50% cost reduction for the bulk of your workload.
Opinionated AI cuts both ways
Opus 5's ability to push back on designs is a leap in reasoning quality — but raises new questions. When should developers listen? When should they override? Balancing openness with independent judgment will be critical with next-gen models.
The LLM pricing war isn't over
Anthropic keeping Opus 5 at $5/$25 while doubling performance signals that competitive pressure remains intense. DeepSeek, Qwen, and Kimi keep pushing prices down. GPT-5.6 Sol is also at $5/$25. Developers are the biggest winners.
Takeaways
- Opus 5 = budget Fable 5. Near-identical benchmarks at half the cost — ideal for most coding and analysis.
- No price hike. Still $5/$25 per million tokens — a free upgrade from Opus 4.8.
- Self-verification is the new skill. The model tests, debugs, and pushes back on flawed designs.
- 26% fewer tokens at max reasoning. Real costs drop below list price.
- Already live on Claude Pro and API. No action needed to start using it.
Content assisted by AI (Amy 🌸). Reviewed by the author.
Related Posts
Better Models, Worse Tool Calling: Claude Opus 4.8 & Sonnet 5's Hidden Regression
Anthropic's newest models produce malformed tool calls ~20% of the time — older models don't. Armin Ronacher's deep dive into why.
Claude Sonnet 5 Launches: Near-Opus Performance at a Fraction of the Cost
Anthropic launched Claude Sonnet 5 on June 30, 2026 — its most agentic Sonnet model yet, with performance approaching Opus 4.8 at a fraction of the price.
Claude Opus 4.8 Is Here: Dynamic Workflows, Effort Control, and a Major Quality Leap
Anthropic just dropped Claude Opus 4.8 with Dynamic Workflows for Claude Code, effort controls, and notable improvements in reliability and honesty. I've tested it — here's what matters.