The Default Model Now Rotates Weekly — Claude Code v2.1.284 Makes Sonnet 5.5 Standard and Hands Admins a Gatekeeper for 'Which Models, and Up to What Cost'
Claude Code's default model was updated twice in one week. v2.1.284 makes Sonnet 5.5 the standard, while mechanisms like deniedModels and a dollar-denominated /usage put 'which models, and up to how much' in administrators' hands.
The default model for "Claude Code," the coding agent that runs in your terminal, has been swapped out twice in the span of just one week. The September 22 release, v2.1.280, switched the Opus-tier standard to the new "Claude Opus 5.5," and the September 28 release, v2.1.284, replaced the Sonnet-tier standard with "Claude Sonnet 5.5." The trend of updating the "brain" you delegate to your agent on a short cycle has now become clearly visible.
The "Standard Brain" That Moved Twice in One Week
Claude Code models have a "default" defined for each tier (Opus / Sonnet). With these updates, that default was replaced with a new generation for both tiers. Both can handle a context of 1 million tokens — enough capacity to hand over long design documents or logs all at once. The pricing differences are as follows. All figures are based on the official release notes.
| Tier Standard Model | Version made default / Date | Context | Input / Output (per 1M tokens) | Cache Read |
|---|---|---|---|---|
| Claude Opus 5.5 | v2.1.280 / Sep 22 | 1M tokens | $4 / $20 | $0.20 |
| Claude Sonnet 5.5 | v2.1.284 / Sep 28 | 1M tokens | $2 / $10 | $0.20 |
In tandem, the September 24 release, v2.1.282, raised the default model for the Pro and Team Standard plans from Sonnet to Opus, bringing them in line with Max, Team Premium, and Enterprise. An adjustment that shifts higher-end models toward the standard is progressing alongside these changes.
New Knobs for Controlling "Which Models, and Up to What Cost"
When models change frequently, organizations become more prone to mix-ups like "the standard changed without us noticing" or "a more expensive model than we expected was running." This series of updates lines up tools that let administrators rein in that uncertainty. Let's organize them by purpose.
Don't Let New Models Get Used Without Permission
v2.1.283 added the admin setting availableModelsMatch. Setting its value to "exact" blocks newly released models from being used until you explicitly add them to a list. It also added deniedModels, which lets you forbid specific models by name, so a policy of "only run vetted models internally" can now be expressed through configuration.
Show Overuse in Dollar Terms
v2.1.284 added dollar amounts to the limit display in /usage, showing this month's spending alongside your cap, as in "$271.40 / $500.00 spent this month." For claude.ai subscription users, a /rate-limit-options command was also added to let you choose how rate limits are handled. "Overuse," which was previously shown as a percentage or remaining balance, has moved closer to something you can grasp in actual dollars.
Audit Your Old Settings
Also introduced in v2.1.283, /doctor prompt-audit inspects instruction files like CLAUDE.md and the way your prompts are written, flagging any old habits left over from previous models. It serves as an entry point for revisiting instructions written on the assumption of the old standard model and aligning them with the new one.
Shoring Up Unattended Operation and Connectivity
Behind-the-scenes stabilization also moved forward. v2.1.284 changed behavior so that a broken response stream now triggers a "retry" instead of displaying a raw error, and it fixed a bug where "Prompt is too long" would linger even after compacting, as well as an issue where MCP tool calls in a resumed session would fail with "No such tool available." The problem of not being able to select a 1-million-token context in Claude Desktop over a gateway was also resolved, with the gateway now automatically detecting supported models. These are quiet but important fixes that pay off the longer you run unattended.
What to Keep in Mind Behind the "Fast Updates"
Having the default model swapped out on a short cycle offers the benefit of immediately being able to use a smarter, faster brain — but it also raises several concerns. First, since output tendencies and costs can shift from version to version, the same prompt is more likely to produce fluctuating results and bills. Second, if the default changes silently, there's a risk that an unvetted model ends up running your production automation. It's natural to read the newly assembled deniedModels, availableModelsMatch, and the dollar-denominated /usage as safeguards prepared precisely against this "changing without permission / overusing." For organizations adopting Claude Code, the safe approach is not to leave everything to the standard's auto-following, but to explicitly pin the models you use and your caps, and to open up new models only after they've been vetted.
References: Claude Code changelog (official) / anthropics/claude-code CHANGELOG.md / Releasebot: Claude Code updates
