Segmind Release Notes (August 31–September 6, 2026)

1. New Models — Sarvam Bulbul v3, Pruna Video Edit, MiniMax H3 Max Turbo

• What's new: sarvam-bulbul-v3-tts brings Sarvam's Bulbul v3 text-to-speech to the API with 37 voices across 11 Indian languages, and a language_code parameter that pins the output language so mixed-script text no longer drifts. p-video-edit (Pruna) is an instruction-driven video-to-video editor that takes a source clip of up to 15 seconds plus up to four reference images, with a draft tier for quick, low-cost previews. minimax-h3-max-turbo is a single slug that runs MiniMax H3 Max Turbo as either text-to-video or image-to-video depending on the inputs you send.
• Impact: Natural Indian-language speech that stays in the language you asked for, prompt-based edits of existing footage, and MiniMax's latest video model behind one endpoint.
• Released: September 3–5, 2026

2. LLMs — Structured Output and Broader Provider Coverage

• What's new: Every LLM on the platform now honours an OpenAI-style response_format, so JSON-schema-constrained requests take effect instead of being silently ignored. The playground gained a structured-output control and the API docs describe the parameter. Anthropic models return a clear explanation when json_object is not supported and extract structured content safely across refusals and truncated replies. Empty chat requests are rejected with a 400 rather than being run — and billed — with filler tokens.
• Impact: Ask any chat model for schema-shaped JSON from the API or the playground and get it, with no charges for empty requests.
• Released: September 4–6, 2026

3. API Keys — Names, Per-Key Model Access, and One-Time Reveal

• What's new: API keys can be named at creation and renamed later (PUT /auth/rename-api-key, or an optional name when generating), and those names now appear in cost analytics, API health, and CSV exports in place of masked strings. A key can be scoped to the models it is allowed to call via PUT /auth/api-key-model-access or the shield dialog on /api-keys, so a leaked credential can only reach the models you chose. The full key is now shown once at creation with a copy button, and is no longer retrievable from the key list. If key creation fails, the console shows the actual reason.
• Impact: Label machine keys by service, limit the blast radius of any single credential, and keep secrets out of the console after the moment you copy them.
• Released: August 31–September 3, 2026

4. Teams — Developer Role and Spending Controls

• What's new: A new Developer team role can create and manage API keys without billing, member, or team-settings powers. Team managers get a Team Spending Cap card on the team page to set daily, weekly, or monthly budgets with live per-period spend (PUT /team/spending-cap on the API), and newly invited members start with a default weekly budget. Budget alert emails now distinguish "nearly used up" (80%) from "reached" (100%), and format amounts to the precision of the limit.
• Impact: Hand out key-management access without granting admin, cap what a whole team can spend, and get alerts that say exactly where the budget stands.
• Released: August 31–September 3, 2026

5. OAuth — Consent Screen, Connected Apps, and MCP Connectors

• What's new: Third-party apps can now complete the OAuth flow end to end: the consent screen at /oauth/consent is live and shows the hostname the authorization code will be sent to, with a plain-language note for local desktop apps. A new /connected-apps page lists every app you've authorized and lets you revoke access. The authorization server also supports Client ID Metadata Documents and floating ports for loopback redirects, which is what tools like Claude Code and the OpenAI connector need to connect to the Segmind MCP server.
• Impact: Connect Segmind to MCP-capable assistants and other apps with a real consent step, then see and revoke those grants at any time.
• Released: September 4–6, 2026

6. Playground and Model Pages

• What's new: Each model's sample inputs are now a visible Try an example strip with thumbnails, and your pick survives a reload. Switching modes in the playground keeps your uploads (Seedance 2.5 no longer drops the start frame or reference media). Image attachments in LLM chat upload reliably on Gemini and other text-and-image models. For ElevenLabs models, the voice_id field searches the full Voice Library with sample playback, and any pasted id resolves to its voice. Long-context token pricing tiers are shown on every pricing surface and returned by the pricing API, and model prices on /pricing, model Pricing tabs, and llms.txt now match what you're billed. Brand searches in the site nav resolve to partner pages, the Stability AI partner page and several others load correctly, and missing or mismatched provider logos were fixed across the model list, Pixelflow picker, and partner cards.
• Impact: Pick an example, keep your inputs while you explore modes, find the exact voice you want, and see the right price before you run.
• Released: September 1–6, 2026

7. Restricted Models — Hidden by Default

• What's new: Models restricted to specific teams now carry a lock badge and are hidden from model listings and Pixelflow pickers unless you turn on Show restricted models — a toggle that is preserved in the URL, so a filtered view can be shared.
• Impact: Cleaner model lists for everyone, with one click to reveal what your team has been granted.
• Released: August 31–September 1, 2026

8. Pixelflow — Sharing, Assets, Connections, and Output Nodes

• What's new: Sharing a Pixelflow to a team now notifies its members. Add to assets works on every output. Multiple connections from one LLM node into a single node keep the same loop mode, so fan-out wiring is no longer dropped, and a connected string field stays in sync with its template. Output nodes gained three size modes (Fit to node, Fit to screen, 100%), a click-to-expand lightbox for single outputs, and larger header controls; node width stays consistent when changing modes, and a trackpad pinch zooms the canvas rather than the browser page.
• Impact: Share and build workflows with fewer surprises, and inspect outputs at whatever size you need without losing the canvas.
• Released: September 4–6, 2026

9. Request History — API Key Attribution Fully Live

• What's new: The API-key column and Key filter on /generations are now fully backed by the platform, so filtering by key returns only that key's requests. The Source: API filter includes traffic from the Python SDK, which previously fell outside it.
• Impact: Trace every generation to the key and client that produced it, with filters you can trust.
• Released: August 31–September 5, 2026

10. Reliability and Clearer Errors

• What's new: The Kokoro-82M TTS worker was rebuilt with roughly 3–4× faster cold starts and a polling loop that fails fast instead of stalling, resolving the queued-job delays some users saw. Upstream provider account faults (for example, a provider's own credit balance running out) are now detected and escalated on our side and no longer appear in your request history or error messages as if your balance were empty; affected Flux requests were never charged. Sarvam requests carry the provider's request id in an X-Provider-Request-ID response header for support traceability.
• Impact: Faster TTS starts, no misleading "insufficient credits" messages for problems that aren't yours, and a reference id when something needs investigating.
• Released: September 1–4, 2026