Opus 4.6 + Sonnet 4.6: Anthropic's February pair, and what "Fennec" actually shipped as

The leaked codename “Fennec” — surfaced from misconfigured Vertex AI logs in late January as claude-sonnet-5@20260203 — did not, in the end, ship as Sonnet 5. It shipped on February 17, 2026 as Claude Sonnet 4.6, at unchanged Sonnet pricing ($3 / $15 per million tokens), twelve days after Opus 4.6 launched on February 5. The slug on this article still says sonnet-5-opus-4-6-launch because it was filed under the leak-era name; the substance is the February pair.

That naming reset matters less than what actually happened in the two weeks between the launches: Opus 4.6 picked up a quiet post-launch safety fine-tuning pass that degraded its agentic behaviour for working developers, and Sonnet 4.6 then arrived and absorbed most of the daily-driver workload that Opus 4.6 had been intended to own.

What shipped, and when

Anthropic’s published positioning for the pair:

• Opus 4.6 (Feb 5): top-of-line model. Claimed wins on Terminal-Bench 2.0, Humanity’s Last Exam, GDPval-AA (≈144 Elo over GPT-5.2), BrowseComp, and MRCR v2 long-context retrieval (76% vs. 18.5% for Sonnet 4.5). Shipped alongside context compaction (beta), adaptive thinking, agent teams in Claude Code, and Claude in PowerPoint (research preview).
• Sonnet 4.6 (Feb 17): “most capable Sonnet yet,” approaching Opus-class on coding and agents. Anthropic reports users preferred it to Sonnet 4.5 in Claude Code testing ~70% of the time, and preferred it to Opus 4.5 ~59% of the time. 79.6% on SWE-bench Verified, against Opus 4.6’s 80.8%. Same Sonnet price tier. 1M context in beta.

If you read those two posts back to back, the framing is awkward by design: Sonnet 4.6 is being sold as Opus-adjacent at one-fifth the cost, twelve days after the new Opus launched. Anthropic is essentially telling you the pair is intentional — Opus 4.6 for the hardest agentic work and the longest context retrieval, Sonnet 4.6 as the daily driver.

Opus 4.6 launch — the HN reception

The Hacker News launch thread for Opus 4.6 hit 2,346 points and 1,031 comments. The early sentiment was warm but quickly accumulated specific gripes: SWE-bench positioning was scrutinised, the rollout-to-Claude-Code timing was the most common practical question, and a thread of comments noted Opus 4.6 felt chattier and more thinking-heavy than 4.5 even on simple prompts.

Four hours later, a second large thread landed: “We tasked Opus 4.6 using agent teams to build a C Compiler” — 735 points, 738 comments — built around a clean-room 100,000-line C compiler in Rust that Claude wrote unattended over two weeks, capable of building Linux 6.9 across x86/ARM/RISC-V plus QEMU, FFmpeg, SQLite, and Postgres. The comment thread split along predictable lines: genuine awe at the artefact, sharp pushback on the cost framing (“$20,000 at current pricing” was called out as VC-subsidised), and a substantial sub-thread arguing the result is mostly a testament to good tests, not to agent teams as a paradigm.

Matthew Berman’s same-day “Claude Opus 4.6 Is Here” walkthrough covered the early-access vibes — podcast post-production, a small Claude Code game build — without engaging the agentic-coding gripes that were already accumulating on HN. That gap between creator coverage and community reception turned out to be the early signal of what came next.

The nerf story (Opus 4.6 specifically)

By the second week of February, the HN thread “Many people have reported Opus 4.6 is a step back from Opus 4.5” captured what working developers were seeing: 5–10× token consumption for identical tasks, unearned confidence, settings tweaks required to get reasonable behaviour. Anthropic’s initial response — paraphrased — was that 4.6 does “deeper thinking” on hard problems and that users should dial effort from “high” to “medium” on simpler ones.

MindStudio’s reconstruction — corroborated by Anthropic’s own later engineering postmortem — is that this was real, measurable, and not a phantom: a targeted post-launch safety fine-tuning pass had spillover effects on multi-step instruction following in long agentic contexts. The model became more conservative in exactly the scenarios — long chains, ambiguous requests, multi-step autonomy — that Opus 4.6 had been positioned to dominate. The fix did not arrive in February. Working developers spent the rest of the month either tuning around it or routing the affected workloads to Sonnet 4.6 once it shipped.

IndyDevDan’s Feb 9 multi-agent orchestration video lands inside this window. The orchestration patterns — Tmux, agent sandboxes — are unchanged from his late-2025 framing; Opus 4.6 slots into the seat that Opus 4.5 occupied. What the video does not engage is the regression conversation already underway on HN — and Dan’s silence here is, in our read, a function of his patterns being structurally insulated from the failure mode (heavy sandboxing constrains the multi-step ambiguity where Opus 4.6 went wrong).

Sonnet 4.6 launch — different reception

The Hacker News thread for Sonnet 4.6 hit 1,346 points and 1,226 comments — smaller upvotes than Opus 4.6, more comments per upvote, and a meaningfully different texture. The dominant themes:

• Prompt-injection safety concerns, citing Anthropic’s own published evaluation showing 8% successful injection takeovers with safeguards plus extended thinking, and 50% under unbounded adversarial attempts
• Labour displacement anxiety — recurring on every Anthropic launch thread, sharper here because Sonnet 4.6 at Sonnet pricing crosses an economic threshold
• The “last 5%” problem — experienced engineers arguing impressive demos still don’t close the gap to production
• Commoditisation economics — what happens to differentiation when frontier-class coding is $3/$15

What did not dominate the Sonnet 4.6 thread: regression complaints. The new Sonnet behaved as advertised out of the gate.

Theo (t3.gg)‘s post-launch “We need to talk about Sonnet 4.6” lands two days after the launch. His read is the harder one for Anthropic: Sonnet 4.6 is good, but at this point Sonnet is still more expensive than GPT-5.4 for comparable agentic tasks, and the gap between Anthropic’s launch narrative and developers’ day-two experience keeps reopening. The video is not a takedown; it’s the kind of measured pressure that lands when a creator who has been consistently bullish starts asking specific pricing questions.

Creator POV vs HN community read

The reconciliation is sharper than usual on this pair.

Creators (Berman, IndyDevDan early-Feb) read Opus 4.6 as a clean tier upgrade with new agentic capabilities to demo. The first-impression videos were positive and stayed inside the use cases that work.

The HN community caught the regression within days, documented it with token-cost telemetry, and stayed on it until Anthropic acknowledged it. The two threads that did the most work — 46902223 and 47052072 — were not the launch celebrations; they were the working-developer follow-up.

Theo’s late-cycle Sonnet 4.6 video sits in between. He’s not echoing the launch deck, but he’s also not piling on. He’s noting that the new pair is genuinely good and still asking why it costs what it costs.

Our read: the gap between creator coverage and HN documentation on Opus 4.6 was the loudest signal of the cycle. When the most technically-engaged comment threads on the web are systematically more skeptical than the launch-day video coverage, that’s the read to trust. The Sonnet 4.6 launch — quieter creator coverage, more substantive HN engagement — is the healthier release pattern.

What this means in late February 2026

Three working positions, based on two weeks of routing across the pair:

1. Sonnet 4.6 is the new default. Drop-in replacement for both Sonnet 4.5 and most of what Opus 4.5 was doing. The ~70% preference rate over Sonnet 4.5 in Anthropic’s own testing matches what we see in client work — fewer overengineered diffs, more consistent multi-step behaviour.

2. Opus 4.6 is conditional. For the longest-context retrieval and the hardest agentic loops, it still wins. For most coding tasks, the regression-era behaviour means you should test before defaulting to it. Pre-fix, the cost-per-good-output math often tilts back to Sonnet 4.6 or even Opus 4.5.

3. Multi-model routing still wins. GPT-5.3 Codex remains the second daily driver for the agent-loop-heavy workloads where Codex specifically shines. The Chinese open tier (DeepSeek, GLM-5, Qwen) keeps cost pressure on Anthropic’s whole stack. If your routing layer is single-provider, this cycle is the one that makes the case for fixing that.

The honest critique

What this launch pair doesn’t mean:

• Anthropic is not communicating regressions well. The Opus 4.6 post-launch tuning issue was visible to working developers within a week and acknowledged only after sustained HN pressure. That’s the second cycle in a row this pattern has played out.
• Sonnet 4.6 cannibalises Opus 4.6’s positioning. At ~59% preference over Opus 4.5 and 80%+ of Opus 4.6’s SWE-bench, at one-fifth the price, the value math for Opus 4.6 is now narrower than the Feb 5 launch deck implied.
• The “Fennec” leak was directionally right and tactically wrong. The model existed; the capability tier was real; the version number was a guess that didn’t land. The lesson is the usual one — leak-era naming is for entertainment, not for planning.

What is durable: Anthropic continues to ship faster than the competition can absorb, the pair-release cadence (Opus then Sonnet, twelve days apart) appears to be the new shape, and the parallel open-frontier compression from Chinese labs means the pricing pressure isn’t easing. The next Anthropic cycle likely lands in late March or April. Plan routing accordingly.

For broader context on how the agentic-coding tier evolved into this cycle, see the AI breaking the SaaS business model analysis.