models.md

title	Models
description	Which Anthropic model assisted fixes use, how to choose one, and what the budget flag limits.

Models

For someone running uptide fix who wants to choose the model and the budget. Measured results are in the provider evaluation; rates in model pricing.

Choosing a model

fix runs assisted fixes only with the Claude Agent SDK, on Anthropic models, with your own ANTHROPIC_API_KEY from the environment. Never put keys in repository config or command arguments. Without the key or the SDK, fix stops before any clone (fix).

npx uptide fix zod --model claude-sonnet-5-5
UPTIDE_MODEL=claude-sonnet-5-5 uptide fix stripe
npx uptide fix zod --max-cost 1

For the model, precedence is --model → UPTIDE_MODEL → nearest uptide.config.json up to the Git root. The model prints before the fix starts, and spend prints at the end.

The optional config accepts only provider and model strings:

{ "provider": "anthropic", "model": "claude-sonnet-5-5" }

Unknown fields, nested settings and key-like values are rejected, including when flags would override them. Keys are read only from the environment. Another provider (openai, gemini), from the config, UPTIDE_PROVIDER or a lone OPENAI_API_KEY / GEMINI_API_KEY, makes fix stop with exit 2: the SDK runs on Anthropic models only.

Defaults

Checked against official documentation on 2026-10-04. OpenAI and Gemini remain for Uptide's own evaluations (UPTIDE_INTERNAL=1, with the patch executor); fix does not run them.

Provider	Default	Reference
Anthropic	claude-sonnet-5-5 (medium effort)	Sonnet 5.5; selected after the storefront comparison
OpenAI	gpt-6.1-sol	GPT-6.1 Sol, through the Responses API
Gemini	gemini-3.8-flash	Gemini 3.8 Flash

Measured on the public storefront fixture (2026-10-04; USD from returned usage). Each cell lists zod / Stripe; every run has a $1 ceiling:

Provider	Model	Verified	Agent sites	Attempts	Cost USD
Anthropic	claude-sonnet-5-5 (medium)	pass / pass	4 / 3	2 / 2	$0.013244 / $0.021592
OpenAI	gpt-6.1-sol	pass / pass	4 / 3	2 / 3	$0.006806 / $0.018677
Gemini	gemini-3.8-flash	fail / fail	0 / 0	12 / 9	$0.000000 / $0.000000

Gemini's serial reruns failed after HTTP 503/429 and transport errors, despite a successful protocol smoke. Its zero confirmed cost excludes $0.473901 / $0.368667 retained for unreported usage; it is not a claim of zero possible billing. Full trials, reservations, times, exact echoed IDs and the Sonnet 4.6/high/medium comparison are in the provider evaluation. These single fixture trials are not a general model ranking.

The executor evaluation compares the SDK with the earlier patch executor on three public repositories.

--max-cost

--max-cost defaults to $1 per package, including zod and stripe. The default is never raised for a large repository: spend goes up only when you raise the limit.

Before the first agent call, fix estimates what the sites left to the agent will cost: 20k input and 2k output tokens per site at the selected model's listed rates, which is $0.06 a site on the default claude-sonnet-5-5 ($2/$10 per million). These are assumptions, not yet calibrated against stored runs. report.json keeps the estimate (llm.estimate) next to the actual spend (llm.costUsd) so the two can be compared. The run prints the estimate and the limit. When the estimate is above the limit, fix stops before any agent call and exits 2, the same as other preflight stops:

estimated $3.00 for 50 sites; limit $1.00. Run again with --max-cost 3.00.

When the agent stops at the limit mid-run, the summary and the pull request name the estimate, the sites left and the limit that continues.

Action and App runs never stop on the estimate: nobody is there to rerun them. They work within maxCost from uptide.yml, and when sites are left, the summary and the pull request say so: … 12 sites not completed. Raise maxCost in uptide.yml to 3.00 to finish.

The SDK executor gives the package's session the limit and enforces it after each turn, so it can overshoot by one turn.

The summary's Phases row and the PR's run details give wall time per phase (analyze, baseline, install, rules, agent, verify, feedback) and the agent's spend in each.

llm.costUsd is computed from Uptide's price table per model and token class; the SDK's own figure is llm.sdkReportedCostUsd, and --max-cost counts the higher of the two. With the SDK, an unknown model's cost is the SDK's figure, marked unverified; with the patch executor, unknown models use Anthropic's highest listed rates with a warning. See model pricing for rates, token estimation and the scope of accounting.

Privacy

SDK sessions send the finding, enclosing code and compiler error, and can also read and send files throughout the temporary clone and the output of type-checks, lint and tests, to Anthropic. Anthropic's data policy applies. Anonymous telemetry, when enabled, adds only the provider and a public model ID; private/custom model IDs become custom. The full model: privacy.