Anthropic Just Finished the Claude 5 Lineup — And Opus 5 Is the One You'll Actually Use

Hey guys, Mr. Technology here.

Friday afternoon, July 24, 2026, Anthropic dropped Claude Opus 5. That completes the Claude 5 generation: Mythos 5, Fable 5, Sonnet 5, and now Opus 5. The only model left in the old generation is Haiku. For the first time in 2026, every Anthropic tier you'd reach for in production — the cheap one, the mid one, the smart one, the overpowered one — is on the new generation.

And here is the part most of the launch coverage is going to gloss over: Opus 5 is the Claude you'll actually default to. Not because it's the smartest — Fable 5 is. Not because it's the cheapest — Sonnet 5 is. But because it is, by a wide margin, the least annoying one to ship.

Let me give you the numbers, the constraints, and the take.

What You Need to Know

>

- Anthropic released Claude Opus 5 on July 24, 2026 at $5 / $25 per million input/output tokens — the same price as Opus 4.8, half the price of Fable 5. - It scores within ~0.5% of Fable 5 on CursorBench 3.2 at max effort, 3× the next-best model on ARC-AGI 3, and posts new SOTA on Frontier-Bench v0.1 and GDPval-AA. - It is not subject to Fable 5's 30-day data retention policy. It ships with 85% fewer cybersecurity classifier trips than Fable 5. Anthropic made it the new default on Claude Max and the strongest model on Claude Pro. - Context: 1M-token window, 128K max output, adaptive thinking on by default.

Why It Matters

• Price/performance is now Anthropic's best argument. Fable 5 at $10/$50 was the ceiling. Opus 5 at $5/$25 is the same intelligence band at half the cost. For most teams, "do I need Fable 5?" just became the wrong question.
• The privacy and operational complaints that pushed teams off Fable 5 are gone. No 30-day retention, lighter cyber classifiers, automatic fallbacks to Opus 4.8 when a request trips a safety filter — that's the model compliance teams have been asking for since June.
• The lineup is now tierable by use case, not by capability ceiling. Haiku for high-volume cheap work, Sonnet 5 for the 80% of agent loops, Opus 5 for the hard 20%, Fable 5 reserved for the research/cyber/biology work it was actually built for. That is a sane production stack.
• It's already the default in Cursor, Devin, Lovable, JetBrains, Kiro, Zapier, Box, and Helix — the entire ecosystem moved in 72 hours. Your IDE is probably already running on it.

The Price Is the Story

Anthropic did something here that the rest of the frontier isn't doing: they held the price flat while shrinking the capability gap. Opus 4.8 was $5/$25. Opus 5 is $5/$25. The "Fable-vs-Opus" gap on CursorBench went from "Opus 4.8 is clearly behind Fable 5" to "Opus 5 is within half a percent of Fable 5 at max effort, and wins on every cost-adjusted axis below that."

The trendingtopics.eu read: "costs half as much as the [Fable 5]." That's the headline. A 1M-token context window at Opus speed and Opus price is a different operating model than a 1M-token context window at Fable speed and Fable price.

For builders running multi-agent pipelines — planner + coder + reviewer + tester — the cost differential between routing everything through Fable 5 and routing through Opus 5 is now roughly 2× across the whole loop. Multiply that by the number of agent steps in your stack. If you have a 12-step pipeline, you've been paying Fable 5 prices for ten of those steps that don't need frontier intelligence.

What Actually Shipped

The benchmark numbers from Anthropic's launch post are unusually specific for a launch blog. A few worth pulling out:

• Frontier-Bench v0.1: New SOTA. More than 2× Opus 4.8's performance at a lower cost per task.
• CursorBench 3.2 (max effort): Within 0.5% of Fable 5's peak score at half the cost per task.
• ARC-AGI 3: Opus 5's score is 3× the next-best model — that's a real reasoning jump, not a vanity one.
• Zapier AutomationBench: ~1.5× the next-best model for the same cost per task. Passes more tasks at the lowest effort setting than any other model does at any setting.
• OSWorld 2.0 (computer use): Outperforms every other model at any given cost. Beats Fable 5's best result at just over a third of the cost.
• Cursor: "Just under Fable 5 and has many of the same behaviors."
• Devin: "Particular strength on difficult debugging and root-cause analysis tasks."
• Box: 8% overall improvement over Opus 4.8, 11% on data analysis, 17% on due diligence workflows.

The life-sciences numbers are the quietly important ones if you ship bio orchem tooling. Opus 5 is 10.2 percentage points higher than Opus 4.8 on inferring molecular structures from spectroscopy data, 7.7 points higher on protein variant function prediction. On every life-sciences eval Anthropic runs, Opus 5 beats Opus 4.8. That is the model for the bench scientists now.

The Real Win: Less Restrictive, Not Less Safe

This is the angle every builder should care about. Fable 5's June launch came with two frictions that broke workflows: the 30-day data retention policy (privacy teams hated it) and aggressive cyber classifiers that intercepted legitimate penetration-testing and vulnerability-research work.

Opus 5 keeps both of those off your back.

• No 30-day retention. Anthropic explicitly carves Opus 5 out of the Fable/Mythos retention regime. If your enterprise security review was the reason you stayed on Sonnet 5, that review just got easier.
• 85% fewer classifier trips. Source: Anthropic's own post. The classifiers are still there — they still block binary-based vulnerability scanning, exploit generation, and penetration testing — but for the 90% of dev work that isn't offensive cyber, you're not getting bounced.
• Automatic Fallbacks (new beta). If a request does trip a classifier, the request falls back to Opus 4.8 instead of returning an error. You get a usable response, not a refusal. This is the single biggest UX improvement in the Claude 5 generation.

The system card also notes that Opus 5 is Anthropic's most aligned model to date — a 2.3 on the automated behavioral audit (lowest of any recent Claude). So the "less restrictive" story is not "less safe." It's "less in your way."

Where It Doesn't Win

Fable 5 still has the lead on the things it was specifically trained for: biology uplift, cybersecurity offense, long-horizon reasoning at the absolute ceiling. If you're running a red team, doing frontier cyber research, or working on protein design at the absolute frontier, Fable 5 is still your tool. Opus 5's safety eval is structured specifically to keep it behind Mythos 5 (the restricted-access variant of Fable 5) on cyber exploit development — that's intentional, that's good, and that's why both models still ship.

The other honest limitation: Opus 5 isn't open-weights, doesn't run locally, and has no announced on-device path. For the open-weights folks building on Qwen, GLM, DeepSeek, or Llama, this is irrelevant. For everyone shipping production agents on a hosted API, it's the new default.

The Take

Anthropic shipped Opus 5 to be the workhorse, not the crown jewel. Fable 5 is the crown jewel. Opus 5 is what you build the production stack on.

If you're a builder choosing a default Claude tier today, the answer is Opus 5 for the hard stuff, Sonnet 5 for the rest. Fable 5 is now a research-tier tool you reach for when you have a specific reason. That's a healthier lineup than Anthropic has had in two years, and it's the first 2026 release where the price held flat while the capability moved up.

The Claude 5 generation is essentially done. Haiku is the last missing piece, and I'd bet on a Haiku 5 announcement before mid-August. After that, the question stops being "which generation of Claude do I use?" and becomes "how do I split work across Opus 5 and Sonnet 5 cost-effectively?" That's a good problem to have.

— Mr. Technology

Quick Summary

• Model: Claude Opus 5
• Released: July 24, 2026
• Price: $5 / $25 per million input/output tokens (same as Opus 4.8, half of Fable 5)
• Context: 1M tokens, 128K max output, adaptive thinking on by default
• Benchmarks: New SOTA on Frontier-Bench v0.1 and GDPval-AA; within 0.5% of Fable 5 on CursorBench 3.2 at half the cost; 3× the next-best model on ARC-AGI 3; 1.5× the next-best on Zapier AutomationBench
• Data retention: Not subject to Fable 5's 30-day policy
• Cyber classifiers: Engage 85% less often than Fable 5
• New feature: Automatic Fallbacks — failed safety checks route to Opus 4.8 instead of erroring
• Default on: Claude Max, Claude Pro (strongest model)
• Adopted by: Cursor, Devin, Lovable, JetBrains, Kiro, Zapier, Box, Helix in the first 72 hours

Sources

• Introducing Claude Opus 5 — Anthropic — primary launch post with benchmark data, customer quotes, alignment numbers, and pricing.
• Anthropic launches Opus 5 — TechCrunch, July 24, 2026 — the same-day press write-up with the 30-day retention carve-out and the Automatic Fallbacks beta confirmation.
• Claude Timeline: Model Release Dates and Key Milestones — independent timeline confirming the July 24 launch, the 1M-context / 128K-output specs, and adaptive-thinking-on-by-default.
• Release of Claude 5 imminent: Anthropic aims to score with lower inference costs — Trending Topics — the "half as much as the Fable" pricing read and the lineup-completion framing.
• LLM Release Timeline — LLM Gateway — confirms Opus 5 is the most recent Anthropic release and lists the parallel July launches (STT 1.0 from xAI, Flash from Google) for context.
• AI Release Tracker — aireleasetracker.com — independent tracker noting Claude Opus 5 as the most recent frontier-model release as of July 27, 2026.
• Introducing Claude Sonnet 5 — Anthropic — the mid-tier counterpart released June 30, 2026, for context on the Claude 5 generation's pricing and positioning.