Simple, sovereign pricing
Invite-only preview today. Team and Enterprise Dedicated as we scale. Usage is metered in credits per developer — and every plan includes the independent referee.
Pilot
- Full agentic workflow in VS Code
- Independent verification (referee) built in
- Data plane in India
- One transferable invite for a friend
Team
- Per-dev keys & metering
- Pooled capacity, fair-use
- Org console + limits
- SSO
Enterprise Dedicated
- Dedicated GPU capacity — no sharing
- In your VPC or air-gapped
- Independent verification + audit trail
- Uptime monitoring + SLA
- Up to 50 developers
Max add-on
- 160K context window
- Dedicated 4-GPU node
- Billed per GPU-hour
- For the largest codebases
Pricing is finalised with each customer in ₹ (INR-first, GST-compliant). Usage metered in credits (tokens × model weight). Dedicated capacity — your own reserved GPU serve, in-VPC or air-gapped — is the flagship enterprise offering: talk to us.
Taksha vs. foreign coding clouds
Same IDE workflow. One difference that matters to a regulated buyer: where the code goes, and who checks the claims.
| Foreign coding clouds | Taksha | |
|---|---|---|
| Where your source code goes | US / foreign cloud | India — or never leaves your VPC |
| Model ownership | Rented frontier API | Owned & operated model family |
| Independent verification of results | Model grades its own work | Referee stamp — a check the model never sees |
| Publishes failure benchmarks | No | Yes — including cells we lose |
| Fine-tuned on your estate | No | Yes (enterprise) |
| Air-gapped deployment | — | Yes |
The Taksha family — pick the right tool
Every model is served on our own sovereign infrastructure, in your region. Heavier models cost more credits (credits = tokens × the model’s weight), so reach for the lightest one that does the job.
Pro flagship
- ▹Full agentic loop: edits, runs, verifies, finishes
- ▹Feature work, refactors, debugging, review
- ▹Any OpenAI-compatible editor or SDK
- ▹45/45 on our graded agentic battery — zero phantom completions
Max Enterprise · on-demand
- ▹45/45 on our graded agentic battery — zero phantoms
- ▹The largest codebases and long multi-file context
- ▹Hardest architectural problems
- ▹Provisioned on request in about 15 minutes
Pariksha
- ▹Unit · integration · functional (pytest, JUnit, Jest, NUnit, Go)
- ▹E2E scripting: Playwright, Selenium, Appium
- ▹Performance scripts: k6, JMeter, Gatling with stated thresholds
- ▹Mutation-checked: it breaks your code to prove the suite goes red
Kriti
- ▹Components that are wired, not just written
- ▹React · Angular · Vue · vanilla
- ▹Graded against a real DOM, not eyeballed
- ▹Exact ids, real event wiring, honest starting state
Yantra coming soon
- ▹Agent frameworks: LangChain/LangGraph, OpenAI + Google agent SDKs
- ▹Tool protocols: MCP servers and clients
- ▹Retrieval pipelines, eval harnesses, guardrails
- ▹For teams whose product IS the agent
Adhunik
- ▹COBOL · VB.NET · mainframe estates
- ▹Proven: COBOL→Spring with exact-arithmetic verification
- ▹VB.NET→C# triple-verified builds
- ▹Referee-stamped at every step
taksha-auto — routed to the model that measures best on your task class
Can't decide? Don't. Each request is classified and routed by a public per-class table — every cell is pass/n from execution-graded battery ledgers, never hand-edited — and every response carries x-taksha-router-receipt saying what routed where, and why. taksha-auto-quality and taksha-auto-cost tune the trade-off.
Any OpenAI-compatible client (Continue, Cline, Zed, Cursor custom base URL). Set the base URL + your key, then pick a model by its id — the proxy meters credits and enforces your plan automatically.
Response headers X-Credits-Used / X-Credits-Remaining show your balance on every call. Locked models return a clear upgrade hint.
Region-pinned. In-VPC. Owned model layer.
Every completion is served from a model we own, on hardware in your jurisdiction. For regulated buyers, that's the difference between a policy promise and a network fact.