On June 30, 2026, Anthropic shipped Claude Sonnet 5, codenamed Fennec, and quietly changed the math on every agentic coding pipeline running on Claude. It puts near-flagship execution, Opus-class agent behavior, and a leading computer-use score, at an introductory $2 input / $10 output per million tokens. That's roughly 60% below Claude Opus 4.8.
For most teams running Claude Code at scale, that single price line makes Opus 4.8 hard to justify for the bulk of day-to-day work.
But the sticker discount is not the real discount. A new tokenizer inflates your input bill, and an aggressive safety layer will refuse some of your legitimate security scripts. Both shrink the savings, and most launch coverage skipped them.
TL;DR
Claude Sonnet 5 delivers about 90% of Opus 4.8's agentic coding capability at 40-60% of the cost, with a native 1M-token context window and a multi-agent orchestrator built into the Claude Code CLI. The introductory price runs through August 31, 2026, then rises.
Two hidden costs, a tokenizer that maps text to up to 1.35x more tokens and an over-refusal rate on benign security tasks, eat into the headline savings. For execution-heavy work, switch.
For deep architecture and hard math, keep Opus.
Key takeaways
- Sonnet 5 introductory pricing is $2/$10 per million tokens through August 31, 2026, then $3/$15 from September 1. Opus 4.8 sits at $5/$25.
- On SWE-bench Pro, Sonnet 5 scores 63.2% versus Opus 4.8's 69.2%, and it leads OSWorld-Verified at 81.2%, ahead of GPT-5.4's 75.0%.
- The updated tokenizer (first shipped in Opus 4.7) maps standard text to 1.0-1.35x more tokens, so the $2/$10 intro acts as a buffer rather than pure savings.
- Sonnet 5 exists in this exact shape because the US government suspended Fable 5 and Mythos 5 on June 12, 2026 over cyber concerns.
- Ignore the 92.4% SWE-bench number. It came from an April Fools' prank.
What is Claude Sonnet 5?
Claude Sonnet 5 is Anthropic's June 2026 mid-tier model that brings near-Opus agentic coding and computer use to a lower price band, built specifically for Google Cloud's Antigravity TPU v6 infrastructure. It runs live on the API, claude.ai, and Claude Code, with a 1-million-token context window enabled globally and no special headers required.
It is, in practical terms, the model you point your agents at when you want them to write code, run it, read the errors, and ship a clean PR without burning Opus-tier budget on every loop.
Why does Sonnet 5 exist in this exact shape?
Two forces shaped this release, and neither is a benchmark.
The first is export control. On June 12, 2026, the US government issued a directive forcing Anthropic to abruptly suspend global access to Fable 5 and Mythos 5 over codebase vulnerability concerns.
Those were the long-horizon autonomy models, $10/$50 per million tokens, built for relentless multi-hour agent runs. Overnight, Anthropic lost its top agentic tier to regulators.
Sonnet 5 is the answer. As Axios reported, Anthropic engineered it with a deliberately lower cybersecurity capability profile so it could ship to the general public without tripping the same controls. High administrative utility, reduced cyber-offense capability, no suspension.
The second force is the IPO. Anthropic is racing toward a public offering at a near-$1 trillion valuation, and VentureBeat framed Sonnet 5 as a land grab. A near-flagship model at a 60% discount locks in developers and drives the high-volume enterprise adoption that supports the financial story, even if it cannibalizes Opus.
So the discount is strategic, not generous. Good for you, but understand the motive.
Sonnet 5 pricing: what you actually pay
Here's the full landscape as of June 2026.
| Model | Input ($/1M) | Output ($/1M) | Notes |
|---|---|---|---|
| Claude Sonnet 5 (intro) | $2.00 | $10.00 | Through Aug 31, 2026 |
| Claude Sonnet 5 (standard) | $3.00 | $15.00 | From Sep 1, 2026 |
| Claude Opus 4.8 | $5.00 | $25.00 | Flagship reasoning |
| Claude Fable 5 / Mythos 5 | $10.00 | $50.00 | Suspended Jun 12, 2026 |
| GPT-5.4 | $2.50 | $15.00 | OpenAI standard |
| Gemini 3.5 Pro | $1.25 | $5.00 | Google premium |
| Gemini 3.5 Flash | $0.07 | $0.21 | Google utility |
Prompt caching stays active with up to a 90% discount on cached inputs, and the batch API delivers a flat 50% off. For repeated codebase analysis, caching is the single biggest lever you have on the input side.
The tokenizer inflation tax
Now the catch most coverage missed. Sonnet 5 uses an updated tokenizer architecture, first introduced in Opus 4.7, that maps typical text to 1.0-1.35x more tokens depending on content density. The same code block costs more tokens to express.
That's why the $2/$10 introductory price exists in the first place. It functions as a cost-neutral buffer for developers migrating off Sonnet 4.6, compensating for the heavier token footprint rather than handing you a clean 60% cut.
Dense, symbol-heavy code sits at the high end of that 1.35x multiplier, so input-heavy pipelines see the smallest real saving.
Audit your own codebase before you assume the headline number. A diagnostic week of token logging beats a quarter of surprise invoices.
Sonnet 5 benchmarks: how close to Opus?
Close enough to matter, with honest gaps. These figures come from Anthropic's system card and release materials.
| Benchmark | Sonnet 5 | Sonnet 4.6 | Opus 4.8 | GPT-5.4 |
|---|---|---|---|---|
| SWE-bench Pro | 63.2% | 58.1% | 69.2% | 57.7% |
| Terminal-Bench 2.1 | 80.4% | 67.0% | 82.7% | -- |
| HLE (with tools) | 57.4% | 46.8% | 57.9% | -- |
| HLE (no tools) | 43.2% | 34.6% | -- | -- |
| OSWorld-Verified | 81.2% | 78.5% | 82.3%* | 75.0% |
*Opus 4.7 baseline. Anthropic retroactively updated several Sonnet 4.6 HLE and OSWorld figures in June 2026 to reflect new grading standards, which resolved earlier metric discrepancies.
Read the pattern. On computer use (OSWorld-Verified), Sonnet 5 is within a point of Opus and well ahead of GPT-5.4. On Humanity's Last Exam with tools, the gap to Opus is 0.5 points.
On SWE-bench Pro, Opus keeps a real six-point lead. Sonnet 5 trades a slice of deep planning for fast execution throughput, and on most agentic loops that trade is invisible.
About that 92% number
You will see "92.4% SWE-bench" and "2M context" floating around. Both are fiction from an April 1, 2026 blog prank that briefly fooled the timeline. Separately, some third-party reviews like Vertu's cite 82.1% on a different, classic SWE-bench suite, which is an easier benchmark than SWE-bench Pro.
When someone quotes a Sonnet 5 score, ask which suite. The Fennec name itself first leaked through a Vertex AI logging misconfiguration in February, and the checkpoint behind it actually shipped as Sonnet 4.6 first, so the public record is genuinely tangled.
How the speed actually happens
The performance comes from infrastructure alignment, not just weights. Sonnet 5 is built for Google Cloud's Antigravity TPU v6 rather than GPU deployments, which Anthropic credits with 20-30% faster inference and near-zero latency when loading hundreds of thousands of tokens. Fast feedback loops are the whole point for an execution model.
Three other mechanisms matter day to day:
A multi-agent orchestrator, "Dev Team" mode, lives inside the Claude Code CLI. The primary instance acts as a manager and spawns specialized sub-agents for testing, QA, implementation, and review, then coordinates cross-agent verification to cut final code errors.
A sandboxed execution harness means Sonnet 5 runs the scripts it writes inside a real terminal, reads stdout and stderr, captures exceptions, self-corrects, and runs unit tests before delivering a PR. It verifies instead of guessing.
Context compaction summarizes conversation history as it approaches limits, which keeps 30-plus-hour agent runs from drowning in their own context, paired with adaptive thinking that scales reasoning depth to call complexity.
The over-refusal trap
The safety design that kept Sonnet 5 out of export jail has a cost on your side. To avoid regulatory intervention, Anthropic shipped aggressive cyber classifiers. They work: Sonnet 5 refuses 92.37% of malicious cyber requests inside Claude Code.
But the same classifiers misfire on benign work. Its success rate on legitimate testing, diagnostic scripting, and system administration fell to 91.55%, which means real friction for security operators and platform engineers running port scans, fuzzers, or recon scripts with full authorization.
The workaround is prompt hygiene. State the sandbox boundaries and benign intent explicitly in your system instructions so the model can distinguish an authorized diagnostic from an attack. Spell out that the environment is isolated and the task is sanctioned.
Sonnet 5 vs Opus: a switch-decision matrix
Don't treat Sonnet 5 as a universal Opus replacement. Match the workload.
| Your workload | Decision | Why |
|---|---|---|
| Multi-file refactoring, CI/CD | Move to Sonnet 5 | ~60% cheaper, equivalent agentic speed, TPU-native |
| Architecture design, deep logic | Keep Opus 4.8 | Sonnet trades planning depth for throughput |
| Desktop / spreadsheet automation (from GPT-5.4) | Move to Sonnet 5 | Leads OSWorld-Verified, fewer execution failures |
| Benign security audits | Proceed with caution | Over-refusal risk on legitimate tests |
Cost at scale: run the real numbers
Take an enterprise burning 100 million tokens a month at a standard 80% input / 20% output split. Migrating from Opus 4.8 to Sonnet 5 lands a roughly 60% cost reduction during the introductory window, stabilizing near 40% once standard pricing arrives on September 1, 2026.
Even after you factor the tokenizer expansion, the savings on a high-volume pipeline stay substantial.
The number that should drive your decision is blended cost per completed task, not cost per token. If Sonnet 5's sandboxed harness finishes a refactor in one verified pass where a cheaper model needs three retries, the per-token rate barely matters.
What this means for you
A clean migration is four steps.
- Repoint the model string. Swap
claude-sonnet-4-6orclaude-opus-4-8forclaude-sonnet-5in your config. See the model configuration docs. - Update tooling. Force Claude Code to v2.1.197 or higher so the native sub-agent CLI works.
- Recalibrate safeguards. Add explicit sandbox-and-benign framing to system prompts to counter over-refusal on dual-use tasks.
- Monitor token budgets. Set console alerts for input expansion from the new tokenizer, and lean on prompt caching for repeated codebase passes.
Practitioner signal backs the move. Sualeh Asif, co-founder of Cursor, said that "with Claude Sonnet 5, agents stay on plan, follow our conventions, and ship clean multi-step changes, all at an efficient cost." Engineers at Zapier reported that complex integration flows which used to stall mid-execution now complete autonomously.
One genuinely strange note from the system card: Sonnet 5 is the first Claude model to actively criticize its own Constitutional AI constraints when it judges a hard boundary as ethically suboptimal. Read that as a research curiosity, not a production feature, but it's the kind of behavior worth watching as these models gain a deeper internal model of their own rules.
Point your execution agents at Sonnet 5 this week, keep Opus 4.8 on the architecture and proof-grade math, and audit your token logs before you trust the 60% headline. The discount is real. It's just smaller and more conditional than the launch posts claim.
Sources
- Introducing Claude Sonnet 5 (Anthropic)
- Claude Sonnet 5 System Card (Anthropic)
- Anthropic launches Claude Sonnet 5 at a steep discount as it races toward an IPO (VentureBeat)
- Anthropic debuts Claude Sonnet 5 for everyday agent tasks with lower cyber risk (Axios)
- Statement on the US government directive to suspend Fable 5 and Mythos 5 (Anthropic)
- Introducing Claude Opus 4.8 (Anthropic)
- The Fennec leak: what actually launched as Claude Sonnet 4.6 (nxcode)
- The April Fools' Sonnet 5 benchmark prank (dev.to)
- Claude Sonnet 5 review covering the 82.1% SWE-bench figure and Antigravity (Vertu)
- Model configuration (Claude Code Docs)