v4.46.0

Changes since v4.45.0 (2026-09-30).

Added

• Add --delete-api-key for the explicit or configured default provider, respecting
  the selected config location and preserving OAuth credentials and global fallbacks.

Changed

• Change the built-in OpenAI default model from gpt-6-luna to gpt-6.1-sol.

Fixed

• Keep the interactive shell running after provider API or connection errors,
  printing the error and rolling back the failed conversation turn.

• Do not load RunBashCode on Windows, even when Git Bash or WSL is installed; use RunPowerShellCode instead.

• Run GitHub CLI commands directly with an argument list so Windows executable paths containing spaces are handled correctly.

• Make --login skip the browser flow when ChatGPT OAuth credentials are already stored.

• Refuse --set-api-key for OpenAI while OAuth details are stored; require
  logout first and preserve existing credentials (#154).

• ChatGPT OAuth Responses: send the system prompt via top-level
  instructions instead of a system input item (fixes
  400 {'detail': 'System messages are not allowed'}), with store:false
  and no temperature on that path. Regular API-key behavior unchanged.

v4.45.0

Changes since v4.44.0 (2026-09-23).

Added

• OpenAI provider: add gpt-6.1-sol variant (1.05M input / 128K output, $2 / $0.10 / $10 pricing).

Changed

• Anthropic provider: replace claude-sonnet-5 with claude-sonnet-5-5 as default model (1M input / 128K output, same $2 / $0.20 / $10 pricing).

v4.44.0

Changes since v4.43.0 (2026-09-22).

Changed

• openai: replace gpt-5.6-sol/gpt-5.6-luna with gpt-6-sol
  ($2.00/$0.20 cache hit/$10.00 output per 1M tokens) and gpt-6-luna
  ($0.10/$0.01 cache hit/$0.50 output per 1M tokens); remove
  gpt-5.6-terra; default model is now gpt-6-luna.
• anthropic: replace claude-opus-5 with claude-opus-5-5 (1M context,
  updated rates $4.00/$0.40 cache hit/$20.00 output per 1M tokens).
• ci: pin both workflows to the ubuntu-24.04 runner instead of the
  floating ubuntu-latest label, which migrates to Ubuntu 26.04 starting
  2026-10-19 (actions/runner-images#14748) — release/publish pipelines
  should not track a moving OS baseline; and bump astral-sh/setup-uv
  from v9.0.0 to v10 in release.yaml and documentation.yml (the v10
  breaking change only affects the default enable-cache: auto mode,
  which this repo does not use).

v4.43.0

Changes since v4.42.0 (2026-09-21).

Added

• providers(zai): new glm-5.3-flashx model — the ~200 tokens/s FlashX
  variant of GLM-5.3, with the same limits as glm-5.3-flash
  (1M context / 128K output); pricing (USD per 1M tokens)
  $0.37/$0.075 (cache hit)/$1.25.

Changed

• tests: rewrite the usage-line rendering tests to comply with
  dev-docs/testing.md Rule 6 (numbers over words). test_turn_report.py
  and test_usage_line_style.py now parse the === ... === summary line
  into labeled parts and assert values composed from the source-of-truth
  format_tokens() / format_elapsed() instead of pinning
  "In: 60/65.5k"-style strings; test_input_tokens_info.py drops its
  local replica of the parts-building logic and renders through the real
  janito.ui.usage._display_usage (close #152).
• providers(xiaomi): replace MiMo-V2.5 with the MiMo-V2.6 series
  (mimo-v2.6-pro, mimo-v2.6-flash, mimo-v2.6-pro-ultraspeed);
  the default model is now mimo-v2.6-flash. Limits are unchanged
  (1M context / 128K output); per-model pricing (USD per 1M tokens):
  Flash $0.14/$0.0028 (cache hit)/$0.28, Pro $0.435/$0.0036/$0.87,
  UltraSpeed $4.35/$0.036/$8.7 (close #150).

v4.42.0

Changes since v4.41.0 (2026-09-16).

Changed

• Renamed the user-facing reasoning-depth setting to effort (close #148):
  the CLI flag is now -e, --effort (was --reasoning-effort) and the
  model-scoped config key is effort (was reasoning-effort, e.g.
  janito --set effort=medium). Breaking rename with no migration, following
  the #77 precedent. The API field (reasoning_effort), provider builtins
  (default_reasoning_effort / supported_reasoning_efforts) and the
  /effort shell command are unchanged.
• Default reasoning effort is now medium for every model that supports it
  (openai, google, alibaba, meta); moonshot/deepseek unchanged since they do
  not support medium (close #147).
• Corrected Xiaomi mimo-v2.5 token limits to 1M input / 128K output
  (was 128K / 120K) to match the official MiMo-V2.5 model page (close #149).

v4.41.0

Changes since v4.40.0 (2026-09-10).

Added

• New built-in atria provider (Atria-Dawn-Preview, https://api.atria-asi.ai/v1, Responses default, stateless). Closes #146.

v4.40.0

Changes since v4.39.0 (2026-09-05).

Added

• Built-in system prompt encourages parallel background tasks via
  StartTask when safe (closes #107).
• Search grounding for Meta muse-spark models (closes #131):
  Responses-only web_search builtin tool plus on_web_search_call /
  on_web_search_done observer events (CLI shows Searching the web...;
  web loop emits web_search / sources events). Citation markers are
  intentionally not rendered: the model's cited sources are low quality.

Added

• Per-model disabled-tools config (closes #144): models with native
  server-side search (Alibaba/Qwen, Meta Muse Spark) hide the external
  WebSearch tool by default; user override per provider/model via
  --set disabled-tools=WebSearch (comma-separated or JSON list).

Fixed

• Corrected ARCHITECTURE.md path in AGENTS.md to dev-docs/ARCHITECTURE.md.
• Privilege badge shows EXEC in red (same as WRITE) instead of yellow.

Changed

• Renamed DeepSeek default model deepseek-v4-flash to deepseek-flash
  with updated per-1M-token rates and refreshed pricing verification date.

v4.39.0

What's Changed

• DRAFT: feat(shell): add -C/--continue to resume the previous CLI session by @albertominetti in joaompinto/janito#137

New Contributors

• @albertominetti made their first contribution in joaompinto/janito#137

Full Changelog: joaompinto/janito@v4.38.0...v4.39.0

v4.38.0

Full Changelog: joaompinto/janito@v4.37.0...v4.38.0

v4.37.0

Full Changelog: joaompinto/janito@v4.36.0...v4.37.0