Tale v0.5.93

Highlights

• Image generation follows each image model's own limits. Tale reads what every OpenRouter image model accepts from OpenRouter's image model listing and asks a model for the closest shape it draws: Recraft's models draw no 3:2, so a landscape is 4:3 there, where every such request failed before. A request with more reference images than the model takes, or an edit model's request without a picture to start from, is refused before anything is billed, and the agent is told why. Vector models, which return SVG files that Tale does not store, are no longer offered, and the Seedream 5.0 models, which OpenRouter lists without a context window, now are (#5057).
• Automatic image model selection no longer recommends GPT Image 1, which OpenAI shut down on October 23, 2026. It tries Gemini 2.5 Flash Image, GPT Image 2.5 Flare (the successor OpenAI names), GPT Image 1 Mini until OpenAI shuts it down on December 1, 2026, then FLUX.2 Pro. The OpenAI provider's catalog lists GPT Image 2.5 Flare, GPT Image 2.5 Sunburst and GPT Image 2 in place of GPT Image 1 (#5057).
• Start an image from a picture you attach in chat: asking to "turn the attached photo into a watercolor" now hands the photo to the image model. Before, the assistant saw the picture but not the file it came from, so it could not pass it on (#5057).
• Through OpenRouter, a GPT image model is asked for medium quality, as Tale already asked OpenAI directly, so its price per image no longer depends on the quality the provider picks: on OpenRouter's default, GPT Image 1 Mini cost about four times its medium price (#5057).
• Sign-in moves to better-auth 1.7.7, the release that fixes two advisories published on October 27, 2026: GHSA-965c-763c-88jm (critical) and GHSA-r4xp-prcw-77qf (high). Both need a better-auth plugin Tale does not enable, Magic Link or OAuth Proxy, so no Tale deployment was exposed; the update clears both from dependency audits (#5061).

Upgrade notes

• No migration runs, and the API contract stays at 3.49.0.
• A sign-in through a provider (single sign-on, SAML or another OAuth provider) that someone started on the previous version can fail to complete after the upgrade: they start it again. Signed-in sessions keep working.
• An organization whose image generation is on Automatic and reaches only OpenAI moves from GPT Image 1 Mini to GPT Image 2.5 Flare. OpenAI prices GPT Image 2.5 Flare's image output at USD 30 per million tokens, where GPT Image 1 Mini's is USD 8; pin GPT Image 1 Mini under Settings › Governance › Models to keep it until December 1, 2026.
• A pin on gpt-image-1 through the OpenAI provider, or on a vector model, now shows as unavailable on the Models page, and agents get no image tool until an administrator chooses another model or Automatic. OpenRouter still lists openai/gpt-image-1 and openai/gpt-5-image, and both fail at OpenAI: choose another model if you pinned one of them. A pinned edit model keeps working for requests that start from a picture.
• With an OpenRouter credential, the backend also reads https://openrouter.ai/api/v1/images/models, a public listing that needs no key, with the catalog listings it already reads from openrouter.ai. An egress rule that admits openrouter.ai covers it; when the listing cannot be reached, image requests keep the facts last read, or Tale's previous defaults.

API contract changes

None in this range. The contract stays at 3.49.0: 163 operations.

What's Changed

• fix(deps): take better-auth 1.7.7 for the sign-in advisories by @yannickmonney in #5061
• docs: write the v0.5.93 release notes by @yannickmonney in #5058
• fix(platform): make image generation follow each image model's limits by @yannickmonney in #5057

Full Changelog: v0.5.92...v0.5.93