Anthropic has released Claude Sonnet 5.5, the second model in its 5.5 generation, slotting it between the already-released Opus 5.5 (for complex reasoning) and the forthcoming Haiku 5.5 (for high-volume, cost-sensitive work). The three-tier structure is now explicit: Opus for judgment-heavy work, Sonnet for everyday coding and document production, Haiku for API-scale throughput. This is product segmentation, not a frontier push. The headline number is Terminal-Bench 4.0, an agentic coding evaluation where Sonnet 5.5 scores 70.6% compared to its predecessor's 10.3% — a nearly sevenfold jump that suggests the previous Sonnet was severely undertrained for multi-step coding tasks rather than that Sonnet 5.5 is miraculous. On FrontierCode 1.1, Sonnet 5.5 at High effort scores 46.2%, trailing Opus 5.5's 54.4% but beating GPT-6 Sol's 49.3% at the XHigh tier. On GDPval-AA, a real-world occupational benchmark, Sonnet 5.5 (1844) nearly matches Opus 5.5 (1846) and significantly outpaces GPT-6 Sol (1487). The model also claims a first for Sonnet-tier: beating Pokémon Red from screenshots alone, a parlor trick that nonetheless demonstrates improved visual grounding. The cost story is where Anthropic is actually competing. Sonnet 5.5 is priced identically to Sonnet 5 ($2/M input, $10/M output, $0.20/M cache reads) but uses fewer tokens per task, yielding up to 30% lower effective cost. Output generation is 30%+ faster. At lower effort settings, Sonnet 5.5 beats Sonnet 5's best scores at roughly one-tenth the per-task cost. This is the real product argument: for well-scoped tasks, you get Opus-adjacent quality at a fraction of the compute bill. The safety section reveals a notable development. Sonnet 5.5's cybersecurity capabilities now match Opus 5's — a significant jump that triggered deployment of Opus-grade cyber safeguards on a mid-tier model for the first time. Higher-risk cyber tasks will visibly fall back to the weaker Sonnet 5. Anthropic has also added anti-distillation classifiers to prevent adversaries from extracting Sonnet 5.5's capabilities through mass-account attacks, and expanded "preserved thinking" to lock reasoning traces to the originating account. These measures acknowledge a real threat: as mid-tier models become more capable, they become more attractive targets for capability theft. The endorsement from Epic Games' COO — praising the model's handling of tens of thousands of lines of gameplay architecture code — and Slack's report of better Slackbot performance with 14% fewer output tokens signal enterprise adoption. But both quotes are from launch partners, not independent evaluators. The pattern is familiar: large labs cultivate strategic testimonials from recognizable brands to signal enterprise readiness. What matters structurally is the pricing ladder Anthropic is building. Three tiers with clear cost-performance tradeoffs, adjustable effort levels within each tier, and safety classifiers that scale with capability — this is infrastructure for API-driven revenue, not a research announcement. The model doesn't advance Anthropic's capability frontier (Opus 5.5 already did that). It fills in the product line. Anthropic is now competing directly on developer economics: cost per task, tokens per completion, latency per output. The frontier lab is becoming a cloud pricing operation. The 20-year question is whether this tiered-model pricing structure concentrates AI capability among a handful of providers who control both the models and the safety gates, or whether the competitive pressure drives costs low enough that capability diffuses widely. For now, the anti-distillation measures and cyber fallbacks point toward tighter control. Anthropic is simultaneously making its models more accessible via price cuts and more locked down via safety infrastructure — a tension that will define the industry's trajectory.
