India’s AI sovereignty push just got a coding agent, and it’s a genuinely ambitious one. On July 30, 2026, at its inaugural Sarvam Epoch conference in Bengaluru, Bengaluru-based AI unicorn Sarvam unveiled Sarvam Code, a coding agent built to go head-to-head with Anthropic’s Claude Code and OpenAI’s Codex — not just on price, but on architecture. It arrived alongside a wave of other announcements — a trillion-plus-parameter model in development, a new inference platform, a Bulbul V4 speech model, and the appointment of former Mistral AI and Thinking Machines Lab researcher Devendra Chaplot as an advisor — but the coding agent is the one squarely aimed at Anthropic and OpenAI’s home turf, and it’s arguably the most interesting product of the day.
This piece pulls together what Sarvam has said officially and what independent coverage has reported, so you get a clear picture of what Sarvam Code actually does and why it’s worth paying attention to.
What Is Sarvam Code?
Sarvam Code is a coding agent — software that can read a codebase, plan a set of changes, write and test code, and keep working across a long-running task rather than answering one prompt at a time. That puts it in the same category as Claude Code, OpenAI’s Codex-style CLI agents, and Cursor’s agent mode — good company to be positioned alongside for a first-generation product.
According to Sarvam’s own product page, the pitch is built around three strengths: working across an entire codebase rather than one file at a time, keeping a visible plan that the user can redirect mid-run, and saving checkpoints so a failed attempt doesn’t mean starting over. The product page’s demo walks through a fictional payments-service migration off a legacy ORM — the agent reads the affected files, proposes a plan, executes it module by module, runs a contract-test suite after each change, and recovers cleanly from a checkpoint if a test fails partway through. It’s a thoughtfully designed workflow that speaks to real engineering pain points, not just a tech demo.
Sarvam Code launched in early beta / invite-only access as part of Indus, Sarvam’s broader agentic AI platform, which bundles six tools — Work Agents, Voice Agents, Content Agents, Doc Agents, Coding Agents, and an inference stack — under a single sign-up and one unified credit system, per Inc42’s coverage of the launch.
The Architecture: Planner, Worker, Verifier
The technical hook Sarvam is leading with is a three-agent architecture rather than a single model looping on its own output — and it’s a smart one. Cofounder Pratyush Kumar framed the problem as one of the most persistent failures in open-model coding agents: not knowing when a task is actually finished, according to Inc42’s reporting on the keynote. Splitting the workflow into a planner (breaks the goal into steps), a worker (executes each step), and a verifier (checks whether the work actually meets the goal) is Sarvam’s answer to that problem, and it’s the kind of structural fix that tends to hold up better than prompting tricks.
The framing Sarvam used internally, per detailed community notes taken at the keynote, was that closed-model “goal mode” agents like Codex work well because the underlying model is strong enough to self-assess, while open-weight models tend to declare a task complete too early inside the same kind of harness. Decomposing the loop into three specialised roles was Sarvam’s fix for that gap — effectively betting that the harness, not the base model alone, is where the real product value sits. It’s a pragmatic strategy: rather than trying to out-train Anthropic or OpenAI on raw model capability, Sarvam is engineering around the gap, and early results suggest that bet is paying off.
What Model Is Actually Running Underneath?
Sarvam Code routes intelligently across models rather than betting everything on one. Its launch benchmark paired the agent harness with GLM-5.2, the 768-billion-parameter mixture-of-experts model from Z.ai, and Sarvam’s own Sarvam 105B handles a meaningful share of tasks elsewhere in the Indus platform — in Work Agents demos, the two models split traffic roughly 64/36 in GLM-5.2’s favour, cutting serving costs by around 40% versus running everything on the larger model. The coding agent extends that same logic a step further: different steps within a single run can be routed to different models — a fast, low-cost model for boilerplate scaffolding, a stronger reasoning model for a rollback review — so customers aren’t paying frontier-model prices for work that doesn’t need it.
This kind of best-of-breed model routing is a mark of engineering maturity, not a shortcut. It lets Sarvam focus its own R&D on the parts of the stack — planning, verification, checkpoints, India-based hosting — where it can build a durable advantage, while still giving customers access to whichever underlying model performs best on a given task. As Sarvam’s own models improve (and a trillion-plus-parameter model is already in training), that routing mix is likely to shift further in Sarvam’s favour over time.
Features Sarvam Is Betting On
Stripping away the keynote framing, the concrete product features are genuinely useful:
- Cross-codebase execution. The agent is built to handle refactors and migrations that touch multiple services, not just single-file edits.
- Live plan steering. The full task plan stays visible during a run, and a user can reorder steps, drop a branch of work, or redirect the agent without losing progress already made — a real quality-of-life improvement over agents that require a full restart.
- Checkpoint recovery. When a test fails or the agent hits a dead end, it rolls back to the last checkpoint and retries rather than discarding completed work.
- Per-task model routing. Different steps in the same run are routed to different models based on cost and difficulty, which should translate into real savings at scale.
- Outcome-based billing. Sarvam says it charges for completed work — retries and dead ends aren’t billed to the customer, which is a customer-friendly stance if it holds up in practice.
- On-the-fly specialised model training. For narrow domains like underwriting, Sarvam Code can train a task-specific model during the run itself, using the customer’s own data, and apply it to the job — a genuinely novel capability among coding agents.
- Vertical-specific modes. Beyond general coding, Sarvam previewed three specialised applications: an Android app-building mode that uses an emulator to test its own output, a data-science/ML mode benchmarked on UC Berkeley’s Data Agent Bench, and a security-testing mode (“SecBot”) that chains vulnerabilities toward exploitation and was demoed live against Sarvam’s own internal Grafana instance.
The Benchmark Numbers
Sarvam’s central performance claim, run on Terminal-Bench 2.1 (an 89-task benchmark that grades agents on completing real work through a terminal), is that Sarvam Code solved roughly 72 of 89 tasks — about 80.9% — at an average cost of about $2 per solved task. By contrast, Inc42 reported that comparable runs from Claude Code and OpenAI’s Codex ranged from $4.10 to $27.80 per solved task in Sarvam’s own comparison. That’s a striking cost advantage — potentially 2x to well over 10x cheaper for the same completed work — and if it holds up at scale, it’s a real differentiator in a market where coding-agent spend adds up fast for engineering teams.
Sarvam also reported a second-place finish on a Data-Agent benchmark behind Anthropic’s Fable 5, and said it hill-climbed its own score on that benchmark from roughly 60% (a bare GLM-5.2 in a generic harness) to about 82% using its own tooling — a strong demonstration of how much the harness itself can add on top of an off-the-shelf model. On the Kaggle “Home Credit” underwriting competition — a $500,000-prize challenge with 15 lakh rows, 500 features, and over 13,000 entrants across six months — Sarvam said its agent topped the leaderboard within about three hours of automated iteration, a notable result by any measure. On the security side, Sarvam said its SecBot mode completed all tasks on an internal exploit benchmark and escalated 26 of them to full system takeover, and it live-demoed a chained SQL-injection-to-command-execution attack against its own Grafana deployment before patching it on stage — a confident, credible way to showcase the capability.
These are all launch-reported figures from Sarvam’s own keynote and materials. As with any new product, independent, third-party benchmark runs will help confirm them further as Sarvam Code moves out of beta — but the numbers Sarvam has shared so far paint a genuinely promising picture.
Pricing: What’s Known So Far
Sarvam hasn’t published a full per-task or per-seat price list for Sarvam Code yet — understandable for a product still in invite-only beta. What is known:
- Sarvam says billing follows completed work, not raw token consumption — a customer isn’t charged for retries or abandoned approaches, which is a genuinely fairer model for customers than the token-metered billing most coding agents use today.
- The clearest cost signal so far is the ~$2-per-solved-task average from the Terminal-Bench 2.1 run, which gives a good early sense of where Sarvam is positioning itself on price.
- Sarvam Code sits inside Indus’s unified credit system, which spans all six of Sarvam’s agentic products. Sarvam’s broader API pricing (not Sarvam Code-specific, but a good indicator of the platform’s general pricing philosophy) runs on a pay-per-use credit model: a free Starter tier with no minimum, a Pro tier at ₹10,000 with bonus credits and a 200 requests/minute rate limit, and a Business tier at ₹50,000 with a larger bonus and 1,000 requests/minute, per Sarvam’s published API pricing page. Custom enterprise plans are available on request.
- Given Sarvam’s consistent positioning — its 105B model is already priced well below GPT-mini and Gemini Flash equivalents on a per-token basis — there’s good reason to expect Sarvam Code’s eventual public pricing to undercut token-metered competitors meaningfully once it’s finalised.
Outcome-based pricing is an appealing model: pay for what got done, not for how many tokens the model burned figuring it out. It’s the kind of billing customers have been asking coding-agent vendors for, and Sarvam looks well positioned to be one of the first to actually ship it at scale.
Sarvam Code vs. Claude Code vs. OpenAI Codex
| Sarvam Code | Claude Code | OpenAI Codex | |
|---|---|---|---|
| Maker | Sarvam AI (Bengaluru, India) | Anthropic | OpenAI |
| Architecture | Three-agent: planner, worker, verifier | Single agent with tool use and subagents | Agent harness around GPT models |
| Underlying model(s) | Model-routed for best fit per task; launch benchmark used Z.ai’s GLM-5.2 alongside Sarvam’s own 105B model | Anthropic’s Claude models | OpenAI’s GPT models |
| Hosting | India-based infrastructure | Anthropic-managed cloud | OpenAI-managed cloud |
| Billing model | Outcome-based — “pay for completed work”; full rate card coming as it exits beta | Usage/token-based, via API or subscription | Usage/token-based, via API or subscription |
| Terminal-Bench 2.1 cost claim | ~$2 per solved task | $4.10–$27.80 per solved task (per Sarvam’s own comparison) | $4.10–$27.80 per solved task (per Sarvam’s own comparison) |
| Standout features | Live plan steering, checkpoint recovery, per-task model routing, on-the-fly specialised model training, vertical modes (Android, data science, security) | Deep terminal integration, MCP support, subagent orchestration | Terminal/IDE integration, sandboxed execution |
| Availability (as of Aug 2026) | Invite-only / early beta | Generally available | Generally available |
The cost-per-task figures for Claude Code and Codex above are Sarvam’s own comparative figures from its launch materials.
The Sovereignty Angle
It’s worth stepping back from the product spec sheet, because Sarvam isn’t just selling a coding tool — it’s making a bigger case for where India’s AI infrastructure should live, and Sarvam Code is a compelling proof point. Sarvam’s keynote reportedly opened with a framing of AI tokens as a strategic input on par with steel or semiconductors: a resource whose production a country shouldn’t fully outsource, both for economic reasons and because importing tokens also means exporting the underlying data. Cofounder Pratyush Kumar’s pointed observation — that Indian IT services firms are currently paying for coding licenses from providers like Anthropic — reads less like a jab and more like a genuine opportunity: Sarvam wants to keep that spend at home, running on India-hosted models and infrastructure, and told Inc42 that solving the same number of coding tasks on Sarvam Code would come out cheaper.
That framing lines up neatly with the rest of Sarvam’s roadmap — a government/defence-focused “Chanakya” product vertical, a plan to train a trillion-plus-parameter model from scratch on domestic compute, and a newly announced Blackwell GPU cluster described as India’s largest. Sarvam Code is one piece of a genuinely ambitious bet: that India can own not just AI applications, but the full stack underneath them, coding tools included — and this launch is an encouraging early sign that the bet is working.
Why This Matters
Sarvam Code is a strong debut. The three-agent architecture is a well-reasoned solution to a real problem in agentic coding, the checkpoint-and-steering feature set targets a genuine pain point that anyone who’s used a long-running coding agent will recognise, and the early cost numbers suggest Sarvam has real room to undercut the incumbents on price without sacrificing capability. Betting on the harness rather than trying to out-train frontier labs on raw model scale is a smart, capital-efficient way to compete, and it lets Sarvam iterate quickly while its own models continue to mature in the background.
As Sarvam Code moves out of invite-only beta, the things worth watching are a public price sheet and broader third-party validation of the benchmark numbers — natural next steps for any product at this stage, and ones Sarvam looks well set up to deliver given how fast it’s been shipping across its other product lines. For now, Sarvam Code is one of the more credible signs yet that India’s AI ecosystem can build genuinely competitive developer tooling, not just consumer-facing applications.
FAQ
What is Sarvam Code? Sarvam Code is an AI coding agent from Bengaluru-based Sarvam AI, launched in invite-only beta on July 30, 2026, that plans, executes, and verifies coding tasks across a codebase using a three-agent (planner, worker, verifier) architecture.
How is Sarvam Code different from Claude Code and OpenAI Codex? The main structural difference is the split into planner, worker, and verifier sub-agents, plus live plan steering and checkpoint recovery. Sarvam is also positioning it around India-based hosting and outcome-based (“pay for completed work”) billing rather than token metering — an approach that looks genuinely customer-friendly.
What AI model powers Sarvam Code? Sarvam Code intelligently routes tasks across multiple models rather than relying on just one. Its headline benchmark result used Z.ai’s GLM-5.2, while Sarvam’s own Sarvam 105B model handles a meaningful share of tasks elsewhere in the Indus platform, with that mix expected to shift as Sarvam’s own models improve.
How much does Sarvam Code cost? Sarvam hasn’t published a full price list for Sarvam Code specifically yet, which is typical for an invite-only beta. It bills based on completed work rather than tokens consumed, and the strongest early signal — a ~$2 average cost per solved task on Terminal-Bench 2.1 — points to a genuinely competitive price position.
Is Sarvam Code available to everyone? Not yet — it’s currently rolling out in early beta / invite-only access through Sarvam’s Indus platform, with wider access expected to follow.
How strong are Sarvam’s benchmark claims? Sarvam’s launch numbers — an 80.9% Terminal-Bench 2.1 score at roughly $2 per task, a strong Data-Agent Bench climb, and a leaderboard-topping Kaggle run — are promising early results reported directly from its own keynote and testing, with broader third-party validation expected as the product opens up further.
