Claude Sonnet 5: Near-Opus Performance at a Fraction of the Price

What builders need to know

• Sonnet 5 shipped 30 June 2026 as Anthropic's mid-tier model, and the company's own framing is blunt: performance "close to that of Opus 4.8, but at lower prices."
• It does not fully close the gap. On the harder agentic-coding benchmarks Opus 4.8 still leads — but by single digits, not the wide margins that used to separate a mid-tier model from a flagship.
• The price gap is the real story. Standard Sonnet 5 pricing is 40% cheaper than Opus 4.8 on both input and output tokens, and an introductory rate through 31 August 2026 widens that further.
• A tokenizer change complicates the sticker-price comparison. Sonnet 5's updated tokenizer can use more tokens per unit of input than its predecessor, so teams migrating should measure their own workloads rather than trust the headline rate alone.

None of this is an announcement piece — Anthropic's release notes and the trade coverage that followed on 30 June already did that job well, including our own coverage of Sonnet 5 landing as the new default in Claude Code with its native 1M-token context window. What we have not seen written down plainly is the number that actually changes a budget: how close is "near-Opus", in benchmark points, and what does that mean for a team in Bengaluru or London deciding which tier to route its production agentic traffic through.

How close is “near-Opus”? The benchmark picture

Anthropic's own system cards give a benchmark-by-benchmark answer, and it is more nuanced than a single "40% cheaper, nearly as good" headline suggests. Opus 4.8 still wins on the hardest agentic-coding and mathematical-reasoning evaluations, sometimes by a wide margin. On several other measures, including one knowledge-work benchmark, Sonnet 5 is statistically even with — or per Anthropic's own account, "slightly outperforms" — its more expensive sibling.

Two things stand out. First, on Terminal-Bench 2.1 — a benchmark that scores an agent's ability to complete real tasks in a terminal, arguably the closest proxy to how a coding agent behaves in production — Sonnet 5 does not just close the gap, it overtakes Opus 4.8. That is a meaningful signal for teams whose agentic workload is mostly terminal-and-tool-call work rather than deep multi-file reasoning. Second, on the two SWE-bench cuts, Opus 4.8 keeps a real but no longer dominant lead: three to six points, not the double-digit gaps that used to separate tiers cleanly. On the hardest problem in this table, competition mathematics, Opus 4.8's lead stays wide — a genuine reminder that "close to Opus" is not "equal to Opus" across the board.

The price side: standard rates, the intro window, and a tokenizer catch

The pricing comparison is where Sonnet 5's pitch gets simple again. At standard rates, Sonnet 5 costs $3 per million input tokens and $15 per million output tokens, against Opus 4.8's unchanged $5 and $25 — a 40% reduction on both legs of the bill. Through 31 August 2026, Anthropic is running introductory pricing of $2 input / $10 output, which pushes the discount versus Opus 4.8 even further for any team measuring cost today.

Read that promotional row carefully, because 31 August is a real cliff, not a soft suggestion. The number to plan a budget around is the standard $3/$15 rate, not the launch price — treat the two-month window as a chance to measure real cost-per-task, not as the resting price you build a forecast on.

The buy-vs-build tier calculus for India and UK teams

The practical question most engineering leads actually have is not "which model is better" — it is "which model should carry our default agentic traffic." That is a tiering decision, and Sonnet 5's benchmark position changes it more than any single release has in a while.

The old default-tier logic went roughly: use a cheap, fast model for simple turns, and reach for the flagship whenever a task looks remotely difficult, because the quality gap between tiers was large enough that guessing wrong was expensive. That logic made sense when a mid-tier model trailed the flagship by fifteen or twenty benchmark points. It makes much less sense when the gap on the benchmarks that map most closely to everyday agentic work — terminal and tool-use tasks, general coding — has shrunk to single digits, or reversed outright. A production pipeline in Bengaluru processing customer support tickets with tool calls, or a London fintech running an agent over regulatory documents with heavy terminal and file-system interaction, is very plausibly running exactly the kind of workload where Sonnet 5 is now the more rational default, not the fallback.

The dual-market framing matters here beyond the obvious currency conversion. Indian teams running high request volumes against thin margins — a common shape for consumer fintech, edtech and support-automation products — feel a 40% per-token saving directly in unit economics, because agentic products are usually priced per seat or per outcome rather than passed through 1:1 on token cost. UK teams, particularly regulated ones in financial services or healthcare adjacent work, tend to weigh the benchmark gap more heavily than the price gap for any workflow that touches a compliance-sensitive decision — which is exactly the workload category where Opus 4.8's remaining lead, especially on precision-heavy reasoning, still earns its premium. Neither market should read this release as "switch everything" or "change nothing"; it is a tiering release, and the right response is to re-tier deliberately rather than either extreme.

Where Opus 4.8 still earns its premium

It would be a mistake to read this release as "Opus is now optional." The USAMO 2026 gap — Opus 4.8 at 96.7% against Sonnet 5's 79.5% — is a seventeen-point margin on competition-level mathematics, among the largest of any benchmark either model is measured on. That is a strong signal that anything genuinely reasoning-heavy, with low tolerance for a wrong intermediate step, still belongs on the flagship: complex financial modelling, multi-step scientific or engineering calculations, or agentic workflows where an error compounds silently across several tool calls before a human reviews the output. SWE-bench Pro's six-point gap tells a related story for coding specifically — on the harder, more realistic split of that benchmark, Opus 4.8's lead is real, even if it is no longer the dominant margin it once was.

The sensible reading is a spectrum, not a binary. Opus 4.8's own dynamic workflows and parallel-subagent features were built for exactly the kind of hard, multi-step orchestration where its benchmark lead shows up — that positioning has not changed, Sonnet 5 has simply made the case for reaching for it selectively rather than by default.

A practical re-tiering playbook

For teams deciding what to actually do this week, a few concrete steps:

• Benchmark your own workload, not the published tables. Run your real production tasks — ticket triage, code review, document extraction, whatever carries your volume — against both models and compare quality and cost on your own data before committing.
• Route by task risk, not by habit. Send high-volume, tool-heavy agentic work to Sonnet 5 by default. Reserve Opus 4.8 for maths-heavy reasoning, high-stakes production changes, and anything where a silent error is expensive to catch after the fact.
• Model the September price, not the July price. Log cost-per-task at the $2/$10 introductory rate now, then re-run the same numbers at the $3/$15 standard rate so the 1 September step-up is not a surprise on the invoice.
• Re-measure token counts after migrating off Sonnet 4.6. The updated tokenizer means an old cost model built on Sonnet 4.6 token counts will not transfer cleanly — measure fresh rather than scale an old spreadsheet.
• Keep a documented fallback. If Opus 4.8 remains part of your stack for the hard cases, make the routing logic explicit in code rather than tribal knowledge, so the decision survives a team change.

The bottom line

Sonnet 5 is not a flagship-killer, and Anthropic never pitched it as one — its own language is "close to Opus 4.8," not "equal to it." What it is, is a mid-tier model that has moved close enough on the benchmarks that map to everyday agentic work, at a price low enough, that the old habit of defaulting ambitious workloads straight to a flagship model no longer holds up under its own arithmetic. For cost-sensitive teams in India and the UK running agentic products at volume, the practical move is not choosing one model over the other — it is re-tiering deliberately: Sonnet 5 as the default carrying the bulk of production traffic, Opus 4.8 reserved for the harder, higher-stakes slice of the workload where its remaining benchmark lead is worth paying for. Teams that made that call already say the maths worked in their favour; the ones still routing everything to a flagship out of habit are the ones most likely to be over-provisioning right now.

Primary sources: Anthropic's launch post at anthropic.com/news/claude-sonnet-5 and the accompanying Sonnet 5 system card. Reporting cross-checked against TechCrunch's 30 June coverage. Verify current pricing against Anthropic's own pricing page before committing spend, as promotional windows and standard rates can change.