Claude Opus 5: What It Is and How It Differs From Opus 4.8 and Fable 5

If you opened Claude Code this week and noticed the /model picker showing a name you had not seen before, that was not a bug. On July 24, 2026, Anthropic shipped Claude Opus 5, and it landed in Claude Code the same day through CLI version 2.1.219. It is now the default model on Claude Max and the strongest model available to Claude Pro subscribers.

If you are a PM or a non-technical builder who just wants to know which model to leave selected, here is the short version: Opus 5 costs the same as Opus 4.8, but performs much closer to the pricier Fable 5 tier on coding and agentic work. For most day-to-day building in Claude Code, it is the new sensible default.

What actually changed

Opus 5 keeps Opus 4.8's exact pricing: $5 per million input tokens and $25 per million output tokens. Nothing more expensive shipped here, Anthropic just made the mid-tier model meaningfully smarter without touching the price tag. Fast mode, which runs roughly 2.5x faster at 2x the base price, now applies only to Opus 5 and Opus 4.8. Opus 4.7 was dropped from fast mode entirely. If you want tighter control over how autonomously Claude Code executes without stopping for confirmation, Ultracode mode is the other lever worth knowing about.

Inside Claude Code specifically, a few defaults shifted alongside the model release:

• The /model picker now highlights the newest model by name, rather than staying visually pinned to "Opus" regardless of version.
• The bundled claude-api skill now defaults to Opus 5, with an in-tool migration path documented for teams still targeting Opus 4.8.
• Subagents can now nest up to depth 3 by default, up from a previous cap of depth 1, which matters if you are running delegation chains where one agent spins up sub-agents that spin up further sub-agents.

On Anthropic's own benchmarking, Opus 5 reportedly doubles Opus 4.8's score on Frontier-Bench, and on OSWorld 2.0 (a computer-use benchmark) it is said to match Fable 5's peak score while using roughly a third of the compute budget. Anthropic also points to a notable jump on its internal organic chemistry and protein-sequence evaluations, though that is less relevant if you are shipping product features than if you are doing scientific research.

Opus 5 vs Opus 4.8 vs Fable 5, in practice

Rather watch than read? I put all three models through the same real test: cloning my app into Lovable, one shot each. The full bakeoff is here:

Opus 4.8 remains available, and nothing about it changed. The difference is purely about what you get for the same money if you switch to Opus 5: closer-to-flagship reasoning on coding, debugging, and long-running agentic tasks, at the Opus price point rather than the Fable price point.

Fable 5 is still the strongest model Anthropic ships, and for the hardest problems (architecture decisions with high blast radius, gnarly multi-file refactors, anything you would not want to get wrong) it likely remains worth the premium. If you are also weighing Anthropic against OpenAI's latest release, see our side-by-side comparison of GPT-5.6 and Fable 5 for product managers. But third-party coverage from VentureBeat and CNBC both frame Opus 5's release primarily as a cost story: a model priced at the Opus tier that behaves close to the flagship tier. That is the calculation worth making before you default every session to the most expensive option.

Independent testing outside Anthropic backs this up directionally. Devin, the autonomous coding agent from Cognition, added Opus 5 support across Devin Desktop and CLI the day before Anthropic's own announcement. On Cognition's FrontierCode 1.1 benchmark, Opus 5 scored 63.6, trailing only Fable 5's 64.9, and beating GPT-5.6 Sol, GPT-5.5, and Claude Sonnet 5. Within Devin, reviewers noted the model favored targeted fixes over sweeping refactors, a trait several called safer for production codebases than more aggressive rewrite-happy models.

GitHub Copilot rolled out Opus 5 the same week, embedding it directly into Copilot rather than requiring a separate Anthropic subscription. Early Copilot testing highlighted the model's strength on long-running agentic tasks: making targeted code changes, validating its own work through regression checks, and coordinating across multiple tools within a single task, while reducing unnecessary execution overhead. Availability spans Copilot Pro+, Max, Business, and Enterprise, though Copilot Free and base Pro users do not get access, and Enterprise/Business admins need to enable it manually in org-wide settings before their teams can use it.

The one caveat worth knowing

Reception has been mixed on tone, if not on capability. Lenny's Newsletter's review called Opus 5 "brilliant (but annoying)," describing strong raw capability paired with more conservative guardrail behavior. Independent testers elsewhere described it the same way: a demanding, opinionated model that outperforms rivals in blind comparisons but pushes back more than earlier versions did. If you found earlier Opus models more accommodating, expect Opus 5 to ask more clarifying questions and push back on ambiguous requests more often. That is a deliberate tradeoff, not a bug, and it is the kind of behavior a PM should actually welcome: a model that flags an underspecified request instead of guessing is one that produces fewer surprises downstream.

Claude Opus 5 for Product Managers

For a PM using Claude Code day to day, the practical impact of Opus 5 is less about benchmark scores and more about what becomes affordable to delegate. Tasks that used to feel like they warranted the Fable 5 premium, tracing a tricky bug across several files, running a longer agentic workflow that touches multiple tools, reviewing a spec against an existing codebase, now sit comfortably inside the Opus price tier.

That shift matters for the kind of work PMs increasingly do inside Claude Code themselves: pulling apart a competitor's product to understand how it is built, reviewing a PRD or spec from an engineering feasibility angle using Claude Code's plan mode, or synthesizing discovery notes into a structured brief. None of that requires the most expensive model on the market, it requires a model that reasons carefully and does not need constant hand-holding. If you are still building the muscle memory for this kind of Claude Code workflow, our Claude Code mastery series walks through it end to end. Opus 5 is built for exactly that gap.

The bottom line

Claude Opus 5 is not a new tier above Fable 5, it is Anthropic closing the gap between its mid-tier and flagship models at the mid-tier price. If Claude Max has already switched you over by default, there is nothing to configure. If you are managing model choice manually, or your team is spread across Devin, GitHub Copilot, and Claude Code, Opus 5 is now the reasonable default for coding and agentic work, with Fable 5 reserved for the problems where getting it wrong is genuinely expensive.