Sarvam Code is Sarvam’s entry into the coding-agent race, announced during Sarvam Epoch 2026. The launch summary describes an invite-only product with a command-line and graphical surface, support for long-running tasks, checkpoints that make work recoverable, plan steering and a pricing idea tied to completed work rather than raw model tokens.
The headline benchmark is more provocative: live coverage says a Sarvam Code + GLM-5.2 configuration solved 72 of 89 Terminal-Bench 2.1 tasks—about 80.9%—for roughly $2 per task. Another event post said the same pairing ranked second on a Data-Agent benchmark behind Anthropic’s Fable 5.
Those claims are worth investigating, not copying uncritically. When this article was published, Sarvam had not yet exposed an indexed product page, public client, run logs or evaluation recipe. That makes the launch an ideal example of why readers need to separate model, agent harness, benchmark and commercial product.
TL;DR — what people are asking
| Question | Direct answer |
|---|---|
| What is it? | An invite-only coding agent from Sarvam |
| Interfaces? | Launch coverage describes CLI and GUI access |
| Long-task features? | Checkpoints, plan steering and recoverable execution were highlighted |
| Underlying model? | The circulated benchmark used Z.ai GLM-5.2 |
| Terminal-Bench claim? | 72/89 tasks, or about 80.9%, launch-reported |
| Cost claim? | Roughly $2 per evaluated task, methodology not yet public |
| Pricing model? | Pay for completed work was discussed; contract details are unknown |
| Can you install it? | No public install command was found when checked |
| Verdict? | Interesting harness thesis; wait for logs, access and reproducible evals |
What Sarvam Code appears to be
A coding agent is more than a chat interface that emits code. It reads a repository, plans, edits files, runs commands, observes errors, manages context and decides when work is finished. That surrounding system is the agent harness.
The launch emphasis on checkpoints and steering suggests Sarvam is targeting the failure mode that matters after a demo: the agent gets 45 minutes into a task, takes a wrong turn and forces a human to restart. A useful checkpoint can preserve repository state, decisions and progress. Plan steering gives the user a way to correct direction without discarding the entire trajectory.
That puts Sarvam Code in the same product category as Claude Code, Codex-style terminal agents, Cursor’s agent workflows and the open harnesses covered in explainx.ai’s complete agent-harness guide and top open and closed harness comparison.
The important unknowns are operational:
- Does a checkpoint include Git state, terminal state and a context summary?
- Can users inspect and restore a checkpoint manually?
- Are commands sandboxed by default?
- Which tools and MCP servers can the agent call?
- Does “plan steering” interrupt a running tool or wait for the next step?
- Are GUI and CLI sessions interchangeable?
- Can enterprises enforce repository, network and secret policies?
Until the documentation answers those questions, “checkpoints” is a product direction, not a complete safety model.
The benchmark claim, translated
The launch summary says 72 successes out of 89 tasks. The arithmetic is straightforward:
72 / 89 × 100 = 80.9%
The denominator matches Terminal-Bench 2.1, which retained 89 tasks while fixing issues in 28 tasks from 2.0. The official 2.1 release note explains that the update addressed tests, instructions and environment problems without changing the benchmark’s basic purpose: measure agents completing real work through a terminal.
Terminal-Bench is valuable because tasks are executed in containers and graded by tests. It is still a system benchmark, not a pure model exam. A score reflects at least:
| Layer | Examples of what changes the result |
|---|---|
| Model | GLM-5.2 reasoning, code quality and tool-use behavior |
| Harness | Planning, retries, context compression and command strategy |
| Tools | Shell implementation, patching, search and file-edit behavior |
| Environment | Container image, dependency access and timeouts |
| Budget | Tokens, wall-clock time, attempts and concurrency |
| Grading | Exact Terminal-Bench version and test patches |
This is why the phrase “Sarvam Code scored 80.9%” is incomplete. The claimed system was Sarvam Code + GLM-5.2 under an undisclosed recipe. A different model in the same harness can score differently; a different harness using GLM-5.2 can also score differently.
Research on harness effects supports that caution. The 2026 Claw-SWE-Bench paper found a large swing under the same backbone when adapter design changed, reinforcing the thesis in our agent harness engineering guide: tools and orchestration can move results by dozens of percentage points.
Why GLM-5.2 is plausible—and why that creates a branding question
Z.ai positions GLM-5.2 for long-horizon coding and agentic work. Its model card publishes results across coding benchmarks and describes the evaluation harnesses used for some tests. GLM-5.2 is therefore a credible model choice for a new agent product.
It also creates the most interesting question around the launch: what is distinctly Sarvam about Sarvam Code if the headline result uses a Chinese open-weight model?
There are several reasonable answers:
- The harness may be the product. Sarvam can add planning, tools, checkpoints, deployment and support around interchangeable models.
- Model routing may be a feature. A task could use GLM-5.2 today and Sarvam-105B or another model when it is stronger or cheaper.
- Indian deployment may be the differentiator. Enterprises may care where inference and repositories run more than who trained the base model.
- The launch benchmark may be a temporary configuration while Sarvam trains coding-focused models.
Sarvam’s June funding announcement explicitly said future research would target agentic, coding and cybersecurity use cases. That supports the fourth possibility, but it does not confirm the eventual Sarvam Code model lineup.
The correct takeaway is not that Sarvam “faked” an agent because it used GLM-5.2. Most coding-agent products support or depend on models trained elsewhere. The takeaway is that buyers should evaluate the whole system and ask which component generates the advertised result.
What does “pay per completed work” mean?
Token billing is easy to meter and hard for customers to predict. A fixed outcome price sounds cleaner: pay when the agent completes a refactor, fixes an issue or finishes a task. It aligns the vendor with completion rather than verbosity.
But completed-work pricing requires a grading contract. Consider a task that asks the agent to “add team invitations”:
- Is completion defined by tests, human approval or a merged pull request?
- Who writes the tests?
- What if the tests pass but security is poor?
- Is a partial fix free?
- Who pays for the model tokens consumed by failed attempts?
- Can the user reject an output indefinitely?
- Does the price include review and follow-up fixes?
Terminal-Bench has binary graders, which makes outcome economics look neat in a benchmark. Real repositories rarely have perfect hidden tests. If Sarvam turns completed-work pricing into a public plan, the most important page may be the definition of “complete,” not the headline rate.
At the reported $2 per benchmark task, 89 tasks would cost about $178 for one pass across the set. That is only meaningful if the run budget, retries, timeouts and failure billing are disclosed. A cheap task can hide a large subsidy during an invite-only preview.
The Data-Agent Bench claim needs a source trail
An event post said Sarvam Code + GLM-5.2 ranked second behind Fable 5 on Data-Agent Bench. This article could not locate an official Sarvam evaluation page or a stable benchmark artifact matching the exact claim during publication.
Data-agent tasks can include loading files, cleaning tables, writing analysis code, producing charts and answering questions. Results are especially sensitive to tool access and filesystem behavior. Before using the rank in a sales or investment memo, request:
- The benchmark’s canonical name and version
- Full leaderboard URL
- Number and type of tasks
- Agent and model configuration
- Pass criteria and judge type
- Run logs or submitted artifact
- Date of the result
“Number two” is not useful if readers cannot identify the leaderboard.
How to evaluate Sarvam Code when invites open
Do not start with a synthetic greenfield app. Use five bounded tasks from a real repository.
Recommended trial set
- Bug fix: a reproducible issue with a failing test.
- Small feature: a UI or API change with explicit acceptance criteria.
- Refactor: behavior-preserving modularization across several files.
- Repository question: trace a data flow and cite exact files.
- Operational task: update a dependency and resolve the resulting breakage.
Scorecard
| Metric | Why it matters |
|---|---|
| Task success | Did it meet the acceptance criteria? |
| Human review time | A passing patch can still be expensive to understand |
| Regressions | Did unrelated tests or behavior break? |
| Intervention count | How often did a person rescue the run? |
| Restore quality | Could a checkpoint recover from a bad direction? |
| Cost | Include failed attempts and review time |
| Security | Did it expose secrets or use unsafe commands? |
Run the same tasks with your current agent. If possible, use the same underlying model to isolate the harness. Our guide to reading AI benchmarks includes a reusable checklist for keeping model, task and budget comparisons honest.
Security questions for an enterprise trial
Coding agents touch source code, credentials, build systems and production tooling. A preview should not receive broad permissions because its benchmark is strong.
Ask Sarvam for:
- Data retention and training-use terms
- Region and deployment options
- Secret redaction and environment-variable handling
- Network allowlists
- Command approval modes
- Audit logs for every tool action
- Repository isolation between users
- Dependency and generated-code scanning
- Incident response and model rollback procedures
If the product supports on-premises inference, distinguish the model endpoint from the agent control plane. A local model does not make a cloud-hosted session log local.
Honest limitations at launch
- The product was described as invite-only, with no indexed public install guide found.
- The 72/89, about $2 per task and Data-Agent ranking are launch-reported claims, not independently reproduced by explainx.ai.
- Sarvam had not published full logs, model settings, token budgets, timeouts or retry rules.
- The supported-model matrix was not public; the circulated evaluation used GLM-5.2.
- Completed-work pricing lacked a public definition of completion and dispute handling.
- A benchmark pass rate does not establish repository security, code maintainability or low review cost.
The explainx.ai read
Sarvam Code is strategically more interesting than “another Indian chatbot.” It suggests Sarvam believes the agent harness—and the enterprise deployment around it—can be a product even when the best underlying model comes from another lab. That is a defensible 2026 position.
The checkpoints and steering story targets a real problem. The outcome-pricing story could be genuinely useful if the grading contract is fair. The GLM-5.2 result is plausible enough to investigate.
Now Sarvam needs to ship the evidence layer: public access, a documented security model, supported-model list, benchmark submission, full run logs and pricing definitions. Until then, Sarvam Code is a promising event reveal, not a proven replacement for an established coding-agent workflow.
Related on explainx.ai
- Sarvam Epoch 2026 complete recap
- Bulbul V4: Sarvam’s expressive TTS launch
- Sarvam AI capabilities, models and APIs
- Terminal-Bench 2.0 and agent evaluation
- How to run GLM-5.2 in coding-agent harnesses
- What is an agent harness?
- How to read AI benchmark claims
- Top open and closed agent harnesses
Sources and verification references
- Sarvam Epoch official agenda
- Sarvam’s Series B announcement
- Terminal-Bench 2.1 release note
- Terminal-Bench leaderboard
- Official GLM-5.2 model card
Product access, pricing and benchmark evidence were checked on July 30, 2026. The launch-reported figures may be superseded by Sarvam’s official documentation; verify the live product page and evaluation logs before making a purchase or migration decision.