Claude Sonnet 5 "Fennec": What Anthropic Actually Shipped

Originally published on the SSNTPL blog. Cross-posted here — canonical URL set in the front matter above so search engines credit the original.

Quick answer: Anthropic officially launched Claude Sonnet 5 on June 30, 2026, four months after the "Fennec" leak first surfaced in Google Vertex AI logs. The launched model is Anthropic's most agentic Sonnet yet, with performance close to Opus 4.8 at lower prices. It scores 63.2% on SWE-bench Pro, 81.2% on OSWorld-Verified, and 57.4% on Humanity's Last Exam with tools — all substantial improvements over Sonnet 4.6 — and ships at introductory pricing of $2 / $10 per million input/output tokens through August 31, 2026, moving to $3 / $15 after that.

The rumor cycle the SSNTPL team tracked since May finally has a resolution: the "Fennec" checkpoint that leaked in February became Sonnet 4.6 first, and the real Sonnet 5 shipped four months later — bigger, more agentic, and priced to become the new default model for most engineering workloads.

For devs and teams shipping agentic products on Claude, the practical takeaway is direct: the Sonnet-vs-Opus decision matrix just changed. If you're weighing whether to build this in-house or bring in dedicated help, SSNTPL's AI testing framework is a useful reference for how this gets validated against real production workloads.

Key Takeaways

• Launched: Anthropic's most agentic mid-tier model to date, released June 30, 2026.
• API model name: claude-sonnet-5.
• Positioning: Near-Opus 4.8 performance at Sonnet pricing; default model on Free and Pro plans.
• Pricing: $2 input / $10 output per Mtok (intro through Aug 31, 2026); $3 / $15 after — versus Opus 4.8 at $5 / $25.
• Context window: 1M tokens; max output 128K tokens.
• Availability: Claude API, AWS Bedrock, Google Cloud, Microsoft Foundry, Claude Code, all Claude plans (Free, Pro, Max, Team, Enterprise).
• Key benchmarks vs Sonnet 4.6 / Opus 4.8: SWE-bench Pro 63.2% / 58.1% / 69.2%; OSWorld-Verified 81.2% / 78.5% / —; Terminal-Bench 2.1 80.4% / 67.0% / —.
• Safety: Lower hallucination and sycophancy than Sonnet 4.6; deliberately weaker cyber capabilities than Opus models.
• Leak accuracy: Fennec codename real. The claude-sonnet-5@20260203 checkpoint real. But that checkpoint became Sonnet 4.6 in February, not Sonnet 5. The genuine Sonnet 5 shipped four months later.

What Is Claude Sonnet 5?

Claude Sonnet 5 is Anthropic's mid-tier flagship model, positioned as the replacement for Sonnet 4.6. Anthropic frames it as its "most agentic Sonnet yet" — a model built to plan, use tools like browsers and terminals, and complete long-running autonomous work that previously required Opus-class systems.

The model sits between Haiku 4.5 (cheapest) and Opus 4.8 (most capable) in Anthropic's lineup, and covers a wider cost-performance curve than any previous Sonnet — from light everyday tasks at low effort to near-Opus-grade autonomous work at extra-high effort.

For most Claude users, it is already the model they are talking to — Sonnet 5 became the default model for Free and Pro accounts on launch day.

When Did Claude Sonnet 5 Come Out?

Anthropic released Claude Sonnet 5 on June 30, 2026 — four months after the "Fennec" checkpoint leaked in Google Vertex AI logs in February. It launched simultaneously across all Claude plans and cloud partners.

How Much Does Claude Sonnet 5 Cost?

Sonnet 5 launched with a cost-neutral pricing strategy relative to Sonnet 4.6.

Tier	Input (per Mtok)	Output (per Mtok)	Notes
Sonnet 5 — Introductory (through Aug 31, 2026)	$2	$10	Set to offset an updated tokenizer that maps text to slightly more tokens
Sonnet 5 — Standard (from Sept 1, 2026)	$3	$15	Same headline rate as Sonnet 4.6
Opus 4.8 (reference)	$5	$25	Roughly 60% more expensive per token
Batch API	50% discount	50% discount	Applied to Sonnet 5
Prompt caching	Up to 90% off input	—	Applied to Sonnet 5

Important caveat: Sonnet 5 uses an updated tokenizer, similar to the change introduced with Opus 4.7. The same input text can map to roughly 1.0× to 1.35× more tokens depending on content type. The introductory pricing window is set to keep the transition roughly cost-neutral compared with Sonnet 4.6.

Claude Sonnet 5 Benchmarks vs Sonnet 4.6 vs Opus 4.8

Anthropic published a full benchmark table alongside the launch. Sonnet 5 improves on Sonnet 4.6 across every published category and closes much of the gap to Opus 4.8.

Benchmark	Sonnet 4.6	Sonnet 5	Opus 4.8	What it measures
SWE-bench Pro	58.1%	63.2%	69.2%	Agentic coding — real-world software engineering tasks
OSWorld-Verified	78.5%	81.2%	—	Computer use — desktop automation
Terminal-Bench 2.1	67.0%	80.4%	—	Terminal-native agentic work
Humanity's Last Exam (with tools)	—	57.4%	57.9%	Frontier reasoning under tool use
GDPval-AA v2 (knowledge work)	—	1,618	1,615	Enterprise knowledge tasks — Sonnet 5 slightly edges Opus 4.8

Two things stand out:

• The Terminal-Bench jump from 67.0% to 80.4% is the biggest single-benchmark improvement in Sonnet 5's launch table, and terminal-native agentic work is exactly the surface where Claude Code sessions live — worth knowing if you're building CLI agents or dev-tooling on top of it.
• GDPval-AA v2 is the one benchmark where Sonnet 5 actually beats Opus 4.8 (1,618 vs 1,615), which is a first for a mid-tier Sonnet model on knowledge work.

Anthropic also reports lower hallucination and sycophancy rates and improved prompt-injection resistance versus Sonnet 4.6, though the company did not release specific hallucination figures.

For a deeper technical comparison across Anthropic's flagship line, SSNTPL's Claude Opus 4.5 vs 4.6 vs 4.7 benchmarks comparison breaks down how each generation moved on the same benchmark suite.

Is Claude Sonnet 5 Better Than Opus 4.8?

Not overall — but the gap is much narrower than any Sonnet before it.

Anthropic's own framing is that Sonnet 5's performance is close to Opus 4.8 on many tasks at a lower price, but Opus 4.8 still leads on the hardest reasoning, most demanding agentic work, and highest-accuracy tasks like deep research and subtle judgment calls.

The practical decision framework:

Choose Sonnet 5 when…	Choose Opus 4.8 when…
High-volume workloads where cost per completed task matters	Highest-accuracy, high-stakes reasoning tasks
Agentic coding, tool use, multi-step automation	Legal, medical, or financial deep-research work
Everyday knowledge work and content workflows	Cybersecurity work (Sonnet 5 is deliberately weaker)
Building agents at scale in Claude Code	Tasks where the marginal accuracy gain is worth 2.5× the token cost
Default routing for Free and Pro users	Extra-high-effort tasks where Opus quality still leads at similar cost

One nuance worth flagging for anyone doing model routing in code: at extra-high effort, Sonnet 5 can actually cost more than Opus 4.8 for similar quality. That's a genuine reason to keep Opus 4.8 in your routing table for the hardest tasks — Sonnet 5 doesn't replace Opus, it repositions it.

What Is New in Sonnet 5 vs Sonnet 4.6?

Six substantive changes:

• Stronger agentic reliability. Anthropic frames the release around agentic follow-through, not a single headline benchmark. Early access partners describe it as finishing complex tasks where earlier Sonnets would stall halfway.
• Self-checking behavior. Testers report Sonnet 5 checks its own output without being explicitly asked to.
• Wider cost-performance curve via effort levels. A single model now covers everything from cheap light tasks to near-Opus-grade autonomous work, depending on effort setting.
• Terminal and browser tool use. Substantial gains on Terminal-Bench 2.1 and OSWorld-Verified translate to more reliable long-running Claude Code sessions.
• Safer defaults. Lower hallucination, lower sycophancy, better prompt-injection resistance, cleaner refusals of unsafe requests.
• Updated tokenizer. Same text can consume slightly more tokens (roughly 1.0× to 1.35× depending on content), offset for now by the introductory pricing window.

What This Means for Teams Building on Claude

The commercial implication is bigger than any single benchmark.

For most teams shipping products on Claude, Sonnet 5 shifts the default answer to three common architecture questions:

• "Which model should we route the majority of agent traffic to?" Sonnet 5 for most workloads; Opus 4.8 only for accuracy-critical fallback.
• "How do we control AI cost as agentic workflows scale?" Route by effort level within Sonnet 5 instead of switching between Sonnet and Opus.
• "What is our fallback model if Fable 5-class capabilities are unavailable?" Sonnet 5 is a stronger daily driver than Sonnet 4.6 was during the Claude Fable 5 access suspension, and Opus 4.8 remains the accuracy fallback.

If you're taking this from prototype to production — building agentic workflows that actually hold up under real usage instead of demo conditions — that's the exact gap SSNTPL's custom software development practice is built to close.

Is Claude Sonnet 5 Safe for Agentic Work?

Anthropic's evaluations found Sonnet 5 to be overall safer than Sonnet 4.6 in agentic contexts, with better refusal of malicious requests and stronger prompt-injection resistance.

Key safety notes:

• Lower hallucination and sycophancy than Sonnet 4.6, though Anthropic did not publish specific figures.
• Deliberately much weaker cybersecurity capabilities than Opus models.
• Real-time cyber safeguards enabled by default.
• For serious cyber, red-team, or offensive-security work, Opus remains the preferred model.

For teams building agent infrastructure, this positioning matters: Sonnet 5 is designed to be the default agentic model for most workloads, with cyber-adjacent work deliberately routed elsewhere.

Where Sonnet 5 Fits in the Wider Claude Family

Anthropic's current lineup is now:

Model	Tier	Best for
Haiku 4.5	Small	Highest-volume, lowest-cost tasks
Sonnet 5 — new default	Mid	Agentic coding, tool use, everyday enterprise work
Opus 4.8	Flagship	Highest-accuracy reasoning, cyber, deep research
Fable 5 / Mythos 5	Restricted	Currently suspended

Anthropic's Mythos-class Fable 5 launch was a major moment, but Fable 5 and Mythos 5 access was suspended globally under a US export control directive shortly after. Sonnet 5 arriving weeks later — with big agentic gains and no Mythos-class restrictions — is not coincidental in the market's reading.

How We Got Here: The "Fennec" Leak-to-Launch Timeline

The 90-Second Timeline

• Jan 28 to Feb 2, 2026: Developers find claude-sonnet-5@20260203 in Google Vertex AI error logs. Codename "Fennec" surfaces alongside. Community goes wide.
• Feb 3, 2026: Predicted release date from the leak. Nothing ships.
• Feb 17, 2026: Anthropic releases Claude Sonnet 4.6 — the same capability tier the Fennec leak described, different name. SWE-bench: 79.6%. Pricing: $3/$15 per million tokens.
• April 16, 2026: Anthropic releases Claude Opus 4.7. No Claude 5 announcement.
• June 9, 2026: Anthropic launches Fable 5 (Mythos-class, public).
• June 12, 2026: US export control directive suspends Fable 5 and Mythos 5 globally.
• June 30, 2026: Claude Sonnet 5 officially launches. Most agentic Sonnet yet. Near-Opus 4.8 performance at Sonnet pricing. Available across all plans, Claude Code, API, AWS Bedrock, Google Cloud, and Microsoft Foundry.

What the Fennec Leak Got Right

• Fennec codename. Confirmed as an internal Anthropic codename.
• The claude-sonnet-5@20260203 checkpoint. Real checkpoint, launched as Sonnet 4.6 first, then Sonnet 5 four months later.
• 1M token context window. Sonnet 4.6 already had it; Sonnet 5 keeps it.
• Sonnet-tier pricing. The leak predicted $3/$15; both Sonnet 4.6 and Sonnet 5 (post-intro) landed exactly there.
• Roughly 80%-plus SWE-bench range. Sonnet 4.6 hit 79.6%; Sonnet 5 keeps improving on the newer, harder SWE-bench Pro at 63.2%.
• TPU-optimized inference. Anthropic's TPU partnership with Google is confirmed, and Sonnet 5 is available on Google Cloud from launch.

What the Fennec Leak Got Wrong

• The name. The February checkpoint shipped as Sonnet 4.6, not Sonnet 5. The actual Sonnet 5 arrived four months later.
• The launch date. Feb 3 was the checkpoint timestamp, not a launch date. Sonnet 4.6 launched Feb 17; Sonnet 5 launched June 30.
• The 82%-plus SWE-bench-Verified score. The Fennec community estimate has effectively been superseded by SWE-bench Pro, which is a harder benchmark; the numbers do not map directly.

The Honest Read

The Fennec story was half right for four months and fully right on launch day. The community correctly identified a Sonnet-tier acceleration in Anthropic's roadmap. It just underestimated how long the road was.

Bottom Line

Claude Sonnet 5 is the model the "Fennec" leak was really pointing at — it just took two model releases (Sonnet 4.6 and Opus 4.7), a Fable 5 launch and suspension, and four extra months to actually arrive.

The final answer to "is Sonnet 5 a big deal?" is straightforward: yes for most teams, no for teams already running Opus 4.8 for accuracy-critical work. It is likely to become the new default mid-tier model across the industry's cost-performance conversation for the next three to six months, until Anthropic ships the next Opus or Haiku release.

For engineering leaders, the practical next step is not a re-platform. It is a re-routing exercise — auditing where in your stack Sonnet 4.6, Opus 4.8, and Fable-class capabilities were previously routed, and re-testing those workloads against Sonnet 5 at low, medium, and high effort levels.

Get that routing right, and Sonnet 5 pays for itself on the token savings within the first billing cycle.

Frequently Asked Questions

When did Claude Sonnet 5 come out?
Anthropic released Claude Sonnet 5 on June 30, 2026. It is available across all plans from launch, is the default model for Free and Pro users, and is accessible to Max, Team, and Enterprise users, in Claude Code, and on the Claude Platform via the API model name claude-sonnet-5.

How much does Claude Sonnet 5 cost?
Standard pricing is $3 per million input tokens and $15 per million output tokens, the same as Sonnet 4.6. Through August 31, 2026, there is an introductory rate of $2 input and $10 output per million tokens.

Is Claude Sonnet 5 better than Opus 4.8?
Not overall. Sonnet 5's performance is close to Opus 4.8 on many tasks at a lower price, but Opus 4.8 still leads on the highest-accuracy reasoning, most demanding agentic work, and deep-research use cases.

What are Claude Sonnet 5's benchmark scores?
SWE-bench Pro 63.2%, OSWorld-Verified 81.2%, Terminal-Bench 2.1 80.4%, Humanity's Last Exam (with tools) 57.4%, and GDPval-AA v2 1,618.

What is Claude Sonnet 5's context window?
Sonnet 5 supports a 1 million token context window with a maximum output of 128,000 tokens.

Where can I use Claude Sonnet 5?
Sonnet 5 is available on Claude.ai (web, iOS, and Android — default for Free and Pro), Claude Code, the Claude API (claude-sonnet-5), Amazon Bedrock, Google Cloud, and Microsoft Foundry.

Was "Fennec" Claude Sonnet 5?
Partially. The claude-sonnet-5@20260203 checkpoint leaked in Google Vertex AI logs in February 2026 under the codename "Fennec," but that specific checkpoint shipped as Claude Sonnet 4.6 on Feb 17. The actual Claude Sonnet 5 launched four months later, on June 30, 2026.

Should we migrate from Sonnet 4.6 to Sonnet 5?
For most workloads, yes — Sonnet 5 beats Sonnet 4.6 on every published benchmark at the same headline pricing after August 31 (and lower during the introductory window). The main caution: the updated tokenizer means the same text can consume up to roughly 1.35× more tokens, so re-test cost against your specific workload before committing to a full production cutover.

---

Originally published at ssntpl.com/claude-sonnet-5-fennec-rumors. For more on agentic AI strategy, model benchmarking, and custom software delivery, visit ssntpl.com.