Pricing: the 2.5x gap is constant across every axis
In Anthropic's official pricing table (platform.claude.com/docs/en/about-claude/pricing), the difference between Opus 5 and Sonnet 5 isn't only in base input/output — the same ratio holds in prompt-cache pricing too: cache-read at $0.50 vs $0.20/MTok, 5-minute cache-write at $6.25 vs $2.50/MTok, 1-hour cache-write at $10 vs $4/MTok. So from whichever angle you look, Opus 5 is exactly 2.5x more expensive.
Simplifying this gap to "Sonnet 5 = the cheap version of Opus 5" would be misleading. That's not Anthropic's official positioning: Sonnet 5 delivers near-Opus-4.8 quality at a much lower price, while Opus 5 differentiates on a separate axis — long-horizon agentic durability and self-verification capacity. Skipping your cache strategy (especially in subagent flows with repeated system prompts) when calculating total cost gives a wrong read on the real cost gap — in a cache-heavy workload the ratio can diverge from the base price.
One more tokenizer note: Claude 4.7 and later models produce roughly 30% more tokens for the same text — which can skew a raw $/MTok comparison somewhat from the real per-task cost.
Durability on long-running agentic tasks
Anthropic specifically emphasized this when introducing Opus 5: the model has markedly improved over previous generations at "verifying its own work and iterating carefully until it succeeds." That means not single-turn Q&A, but an agentic task that can sustain the edit-file, run-test, read-error, fix loop for a long stretch. The jump on Frontier-Bench v0.1 and CursorBench (more than 2x Opus 4.8's performance, within 0.5% of Fable 5's peak) backs this claim numerically.
Sonnet 5's announcement has a similar improvement too — but the reference point is Sonnet 4.6, not Opus 5: it includes a customer quote saying it "completes complex tasks where previous Sonnet models stalled." So Sonnet 5 is advancing within its own generation, but there's no direct durability comparison against Opus 5 on record.
In practice this means: if you need to complete a migration or a multi-file refactor in a single long session with minimal human intervention, Opus 5's self-verification capacity reduces the risk. The same task with Sonnet 5 will likely need more frequent checkpoints — a human-time cost that doesn't show up in the pricing table.
Effort flexibility and the feel of latency
Both models share the same adaptive-thinking architecture: 5 effort levels (low→max), with "high" as the default on both. Anthropic's official guide gives a clear recommendation here: start with the default (high), then adjust up or down based on your evals — meaning effort should be your first lever before switching models. This can be read as an official warning against the reflex of "automatically give every hard task to Opus": sometimes running Sonnet 5 at "high" effort gives a better result than running Opus 5 at "low" effort.
The latency difference is clearer: the official model comparison table classifies Sonnet 5 as "Fast" and Opus 5 as "Moderate." This is felt in interactive scenarios where the user is waiting for a response. Opus 5 also offers a research-preview feature called "fast mode" — up to a 2.5x increase in output speed — but it comes with premium pricing, and Sonnet 5 doesn't have this option.
Bottom line: if the work is interactive and the effort setting gives sufficient quality, Sonnet 5's speed advantage is one more reason to prefer it. If the work runs in bulk/in the background (batch, subagent, cron), latency stops mattering and the decision comes down entirely to the quality/price balance.
Fit for subagent delegation
One clarification matters here: Anthropic's official documentation doesn't provide a delegation matrix of "give this task type to this model." The official advice stays at the level of "adjust effort first, switch models if needed." The rule "give discovery/mechanical work to Sonnet, specification-clear complex code to Opus" isn't an official Anthropic recommendation — in practice it's a working discipline that teams using coding agents have developed.
That said, the two models' official feature differences do support this split: Opus 5's self-verification and long-agent durability make it well suited to "complex work with an unclear finish line," while Sonnet 5's speed + low-cost combination makes it well suited to "high-volume work with a clear finish line."
There's a practical trap in the Claude Code CLI: according to the release notes, a subagent's model resolves first from the call's own model value, then from the default subagent setting if that's absent, then from the main session's model if that's absent too. So even a simple file-scanning task where you forgot to specify the model can end up running at the expensive model's price. Explicitly specifying the model on every delegation — "discovery → Sonnet 5", "specification-clear code → Opus 5" — is the safest approach.
Total cost on the same task: why the simple ratio math misleads
In its simplest form, using the 2.5x ratio to say "Opus 5 is 2.5x more expensive than Sonnet 5" is mathematically correct but incomplete in practice. Anthropic's announcement pages show cost-performance charts visually, but they aren't published as a numeric table — so there's no precise official figure for "total cost on the same task."
Three factors affect the real total cost: (1) cache usage ratio — in a flow with a repeated system prompt, cache-read/write ratios behave differently from the base price; (2) number of iterations — if Opus 5's self-verification capacity can finish a task in fewer turns, it can end up cheaper in total even at 2.5x the per-token price; (3) tokenizer difference — Claude 4.7 and later models produce roughly 30% more tokens for the same text.
Practical advice: run a critical workflow on both models with a small sample (same prompt, same effort level) and measure the actual token count and turn count — decide based on a figure measured on your own workload rather than trusting the theoretical 2.5x ratio.
Retirement/support schedule
Per the official model deprecation page, both models are "Active": Opus 5 will be retired "not before July 24, 2027," Sonnet 5 "not before June 30, 2027." The one-month gap between them doesn't create a practical planning difference, but it's a date every team pinning a model to production should know.
The launch dates are close too: Sonnet 5 was announced on June 30, 2026; Opus 5 on July 24, 2026. That 3.5-week gap might make Sonnet 5 look "newer," but the knowledge cutoff dates say the opposite: Opus 5 is May 2026, Sonnet 5 is January 2026 — Opus 5's training data is 4 months more current.
One common misconception needs correcting here: undated IDs like claude-opus-5 and claude-sonnet-5 are not "evergreen pointers that resolve to the latest release." In the official documentation's own words, starting with the 4.6 generation and later, an undated ID is that release's canonical model ID and corresponds to a single, fixed snapshot; Anthropic doesn't update the weights of an existing ID, a new release ships under a new ID. The alias behavior only exists for pre-4.6 models. Don't confuse this with short names like default/opus in the Claude Code CLI — those are a separate layer (see Common Pitfalls).