Anthropic’s Sonnet 5 Just Cannibalized Opus, and That’s Not a Coincidence

Anthropic dropped Claude Sonnet 5 on Tuesday, calling it "the most agentic Sonnet yet" and positioning it as a cheaper alternative to Opus for long-horizon coding and tool use. Multiple outlets (TechCrunch, The Decoder, VentureBeat, Simon Willison) covered the rollout within hours, which tells you everything you need to know about how scared the incumbents in the agentic-coding lane just got.

The framing Anthropic chose is not accidental. "Most agentic" is not a benchmark. It is a category claim, and Anthropic is planting a flag in the exact ground OpenAI Codex and Gemini 3.1 have been fighting over for the last quarter. Sonnet 5 undercutting Opus on price is not generosity. It is a defensive pricing move, and the "agentic" framing is the new battleground. Pricing is the weapon.

What ships, what changes

Sonnet 5 lands as a mid-tier model with explicit positioning for "long-horizon coding and tool use," the two workloads that actually pay Anthropic's bills in 2026. Anthropic is not hiding what this is for. The marketing copy leads with agents, not chat, because that is where the revenue migrated six months ago and where the competitive pressure is now sharpest.

The interesting row is the second one, not the first. Sonnet 5 is not priced against Opus. It is priced against the agentic workloads where Codex and Gemini 3.1 are already winning real deployments. Anthropic is using its own mid-tier as a precision weapon in a specific lane.

A real-world routing change

If you run Claude Code with model routing, Sonnet 5 is the new default for the 70% of tasks that are "multi-step, tool-using, but not particularly clever." That is most production agent traffic. Here is roughly the swap most teams will make this week:

# .claude/agents.yaml — before Sonnet 5
agents:
  coder:
    model: claude-opus-4-8
    max_iterations: 30
  reviewer:
    model: claude-sonnet-4-6
    max_iterations: 10

# .claude/agents.yaml — after Sonnet 5
agents:
  coder:
    model: claude-sonnet-5        # cheaper, "most agentic" - default forward
    max_iterations: 30
  reviewer:
    model: claude-sonnet-5        # same model, different prompt - small win
    max_iterations: 10
  deep:
    model: claude-opus-4-8        # now reserved for the 10% that need it
    max_iterations: 60

Two of three roles drop to Sonnet 5. Opus 4.8 becomes the escape hatch, not the default. That is a meaningful structural shift in how teams pay Anthropic, and it is exactly the move an incumbent makes when it is more afraid of the next wave than its own cannibalization.

Anthropic quietly let Sonnet 4.6 keep its mid-market slot. Sonnet 5 is not replacing 4.6, it is sitting above it. Sonnet 4.6 stays the cheap steady default for high-volume background work.

The defensive read

The "cheaper than Opus" line is being read in two contradictory ways. Some commentators frame it as Anthropic being generous, democratizing agentic capability for teams that could not afford Opus. Others read it as a panic move, a price cut because the alternative was watching Codex and Gemini 3.1 eat the mid-tier while Opus stayed premium-priced and slow.

Both readings understate it. Anthropic is doing the move a hardware company does when it knows a successor is coming and it wants to pre-empt the discount cycle. By undercutting itself on the mid-tier before OpenAI or Google can, Anthropic locks in agentic-coding revenue on multi-year contracts at a price competitors will have to match just to bid. That is not generosity. That is a moat dug in advance, paid for out of margin it was going to lose anyway.

The signal to watch over the next 30 days is whether Codex and Gemini 3.1 cut price in response. If they do, Sonnet 5 will have just defined the new floor of the agentic-coding tier and everyone will be repricing against it. If they do not, Sonnet 5 will have eaten a chunk of their pipeline while they waited to see what happened. Either branch ends the same way: Anthropic's mid-tier wins by default.

Source: latest.json pitch #1 (https://www.techcrunch.com/2026/06/30/anthropic-launches-claude-sonnet-5-as-a-cheaper-way-to-run-agents/), cross-referenced with The Decoder, VentureBeat, and Simon Willison coverage of the launch.