What Anthropic actually shipped on June 30 and why the price tag is the story
Anthropic launched Claude Sonnet 5 on June 30, 2026 as the new default across free, Pro, Max, Team, and Enterprise plans — available inside Claude, Claude Code, and the Platform API on day one. The Sonnet class historically anchored the mid-tier coding-and-general-work substrate at cost-tier price; Sonnet 5 shifts the tier without changing the label. The model runs on a native 1M-token context window, scores 63.2% on the agentic-coding surface (vs. Opus 4.8's 69.2% and Sonnet 4.6's 58.1%), and edges Opus 4.8 on the knowledge-work aggregate. Introductory pricing: $2 per million input tokens and $10 per million output tokens through August 31, then a step-up to $3/$15.
The operationally important reads:
- Sonnet 5 is priced under Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on the agentic-coding surface. The only cheaper frontier substrate for coding-agent workloads is Gemini 3.5 Flash at $1.50/$9. The team routing multi-step coding tasks against Opus-tier pricing without re-running its per-workload shootout against Sonnet 5 is leaving margin on the table for a substrate whose 6-point accuracy gap against Opus 4.8 closes cleanly for verifier-guarded workloads.
- The 1M-token context is native, not extended with rope tricks or windowed retrieval. The context-engineering-replaces-RAG pattern the market shifted to in Q2 2026 grades directly against a native long-context substrate; Sonnet 5 makes the pattern cheaper to run at the default-routing tier, not just at the escalation path.
- Agentic execution is now baseline at every price tier, not a Pro-tier premium. Anthropic's public framing acknowledges the shift: capability is no longer the axis competition runs against — cost-per-successful-task is. The team whose per-prompt routing policy still grades against last cycle's Opus for hard tasks, Sonnet for cheap tasks split is running against an obsolete tier map.
- The promotional pricing window closes on August 31. The two-month discount is the procurement function's window to run the per-workload shootout, ship the routing-policy update, and lock the FY27 standing contract before the price step-up. Teams that defer the shootout to Q4 pay the 50% price jump on workloads that migrated to Sonnet 5 anyway.
The structural read isn't Anthropic shipped a cheaper Sonnet. It is that the coding-agent frontier's per-workload cost curve stepped down a tier at the mid-price point, the 1M-context substrate is now the default-routing tier, and the FY27 model-routing matrix drafted six months ago needs a per-workload-class re-shootout inside the two-month promotional window.
What the Sonnet 5 tier restructures for the FY27 routing matrix
The default-route for coding-agent workloads bifurcates by verifier-coverage-gap, not by workload difficulty. Twelve months ago, the mental model was route hard tasks to Opus and easy tasks to Sonnet. The Sonnet 5 pricing-and-accuracy envelope collapses the split: route verifier-guarded coding workloads to Sonnet 5 as the default, and escalate to Opus 4.8 only when the verifier coverage gap on the specific workload class exceeds the accuracy-gap-times-error-cost budget the team can underwrite. The per-workload verifier is now the routing-decision input, not the per-prompt difficulty heuristic.
The eight-parallel-worktree-agent pattern gets cheaper without a new orchestrator. The pattern Cursor 3 standardized on grades against per-agent cost as much as per-agent latency. Sonnet 5 at $2/$10 through August drops the per-worktree-agent cost by roughly 40% versus Opus 4.8 at $3.75/$15 without changing the coding-loop scaffold. The team's coding-throughput surface improves against the same orchestration substrate — the change is in the routing policy, not the tool chain.
The two-vendor coding-frontier standing contract on the FY27 plan gets a stronger Anthropic-side anchor. The four-vendor frontier map (Anthropic Opus 4.8 / Sonnet 5, OpenAI GPT-5.6 Sol, Google Gemini 3.5 Flash / Gemini 3 Deep Think, DeepSeek V4 open-weights) grades against per-workload-class portability, not vendor loyalty. Sonnet 5 strengthens the Anthropic side of the dual-vendor anchor at the cost tier where GPT-5.5 and Gemini 3.1 Pro previously held the price-per-successful-task lead; the standing contract negotiation this quarter runs against a stronger Anthropic offer than last quarter's.
Safety-and-reliability gains show up in the operational metric, not the marketing claim. Anthropic ships Sonnet 5 with lower refusal-of-benign-requests rates, tighter tool-use compliance, and reduced hallucination on structured-extraction workloads. The load-bearing read is the operational-reliability delta on production coding agents grounded in the team's own verifier suite, not the aggregate safety-benchmark rank. The team that re-runs its per-workload reliability suite against Sonnet 5 this sprint measures the actual delta; the team that reads the marketing claim ships the assumption.
Where the Sonnet 5 launch is signal and where it is noise
Signal: the 63.2% agentic-coding score at the Sonnet-tier price is the real news. The accuracy-per-dollar frontier moved. Every coding-agent workload class whose FY27 routing decision was made at 58% Sonnet 4.6 accuracy is a candidate for re-shootout at 63.2% Sonnet 5 accuracy. The delta compounds across a full quarter of coding-agent runs.
Signal: the promotional pricing window is a hard deadline, not marketing spin. The August 31 step-up is set — the procurement clock the FY27 standing-contract negotiation runs against is the two-month window between now and the price jump. Teams that finish the shootout, ship the routing update, and lock the contract inside the window buy the price at the introductory tier for the annual commit.
Noise: the 6-point accuracy gap versus Opus 4.8 is not a substrate-wide disqualifier. The gap is the escalation-path budget, not the default-route disqualifier. Verifier-guarded workloads close the 6-point delta cleanly; free-form-generation workloads whose failure mode is not verifier-catchable still route to Opus 4.8. The per-workload routing policy is the artifact that encodes the split, not the cross-workload aggregate rank.
Noise: Sonnet 5 replaces Opus 4.8 is the wrong frame. Opus 4.8 stays on the routing matrix as the escalation path. The right frame is Sonnet 5 becomes the default-route substrate at the cost tier where the FY27 procurement plan needs the price-per-successful-task to sit; Opus 4.8 stays the escalation substrate for the verifier-coverage-gap workloads whose accuracy cost the team cannot underwrite at the Sonnet tier.
What the engineering team should do inside the next two weeks
Run the per-workload-class shootout on Sonnet 5 against Sonnet 4.6 and Opus 4.8 inside two weeks. For the team's top-three coding workload classes (multi-file refactor against explicit test contracts, dependency upgrade against explicit version pins, structured extraction against deterministic schema), measure per-class pass-rate, per-class time-to-completion, per-class per-token cost, and per-class verifier-coverage-gap. The output is the routing-policy update artifact the standing-contract negotiation runs against.
Update the per-prompt routing policy to flip Sonnet-tier from cost-tier fallback to default-route. The routing-policy artifact in the team's repo is where the substrate shift lands. Ship the default-route flip against agentic-coding workloads inside the two-month promotional window, and write the Opus-tier escalation path against the verifier-coverage-gap workload classes the shootout identified.
Re-grade the coding-agent throughput surface against the new per-token cost envelope. The per-worktree-agent concurrency cap and per-agent budget the coding-agent loop runs against were set against the prior tier's per-token cost. Re-grade the caps against the ~40% cost drop and ship the updated coding-throughput budget inside the sprint. The team's per-week throughput surface improves without a new orchestrator.
Lock the FY27 standing contract before August 31. The introductory price is the negotiation anchor; the two-month window is the deadline. The procurement function that runs the shootout, ships the routing update, and locks the annual commit inside the window pays the $2/$10 rate on the workloads that migrated to Sonnet 5 anyway. The function that defers pays $3/$15 in September on the same workloads.
What Sonnet 5 makes cheaper but does not replace
Sonnet 5 compresses the per-token cost of the coding-agent surface's default-routing tier, not the senior judgment of deciding which workload classes are Sonnet-tier-shape, writing the verifier the routing policy grades against, owning the per-vendor portability envelope on the standing contract, and running the per-cycle routing-policy code review against the team's coding-agent loop. The teams that confuse the cheapened per-token cost for cheapened judgment route the free-form-generation surface against a substrate whose verifier coverage gap does not close, and read the per-cycle post-mortem on the routing-policy gap the shootout would have surfaced. The teams that keep the senior judgment at the center of the routing decision translate the substrate shift into per-week throughput improvements the prior tier map could not produce.
The model-routing question is no longer which model is the flagship; it is which workload classes the Sonnet-5 substrate is the default-route for, which workload classes the Opus-4.8 substrate is the escalation path for, and which per-vendor portability envelope the FY27 standing contract underwrites against the four-vendor frontier map.
At SONNET CODE we run the AI Development engagement against the per-prompt routing policy artifact — per-workload-class shootouts against the four-vendor frontier map, per-vendor portability envelopes on the FY27 standing contract, and per-cycle substrate-shift code reviews against the team's coding-agent loop. If your team's routing policy is still written against Sonnet 4.6 pricing and accuracy, schedule a call — we'll walk you through the routing-matrix update we ship inside one sprint, well before the August 31 price step-up closes.
