OpenCode vs Claude Code vs OpenClaude: Benchmarks
Only one of these three coding agents shows up on a benchmark leaderboard. The Terminal-Bench 2.1 numbers, the missing scores, and how to choose.
By Capital & Compute
Ask which of these three coding agents wins on the benchmarks and you have already walked into the trick. Of OpenCode, Claude Code, and OpenClaude, exactly one appears on any benchmark leaderboard. As of July 2026, the Terminal-Bench 2.1 leaderboard at tbench.ai lists Claude Code in four of its thirteen entries, topping out at 83.1%. OpenCode, the most-starred open-source coding agent on GitHub, appears zero times. OpenClaude, the Claude Code derivative that has been climbing GitHub since spring, also appears zero times, and its own README claims no benchmark results at all.
That absence is not a footnote. It is the single most important fact in this comparison, because every “OpenCode vs Claude Code” verdict you have read was written without a measured score on either side. The popular comparisons are hands-on impressions: Builder.io’s January 2026 test timed four tasks in Docker containers, and Composio’s June 2026 write-up logged a hundred hours of daily driving. Useful reading. Not benchmarks.
So this comparison does something different. It lays out the one dataset where the harness itself is the measured variable, is honest about where the data runs out, and then compares the three tools on the axes you can actually verify today: cost structure, provider freedom, extensibility, adoption, and legal posture.
Three tools, and the names people mix up
The names in this matchup are a small disaster, so definitions first.
Claude Code is Anthropic’s own coding agent, closed source, tied to Anthropic models, sold inside the Claude Pro ($20 a month), Max ($100 and $200) and Team plans or metered at API rates. It is the harness whose extension system (skills, hooks, subagents, MCP) this site has covered in depth.
OpenCode is the open-source coding agent from the SST team, now under the Anomaly organization on GitHub: MIT licensed, 183k stars and 22.7k forks as of July 7, 2026, with a v1.17.14 release shipped July 6. The software is free. opencode.ai claims support for 75+ LLM providers through Models.dev, local models included, plus automatic LSP loading and a claimed 7.5 million monthly developers (a vendor figure; the site also still displays a stale 160k star count, so read its marketing numbers with that in mind).
OpenClaude is the odd one out. The Gitlawb/openclaude repository describes an open-source coding agent CLI for cloud and local providers, 29.8k stars and 8.9k forks as of July 7, 2026, release v0.22.0 shipped July 6. Unlike OpenCode, it is not a from-scratch project: the README states it originated from the Claude Code codebase and has since been substantially modified. Its license section says MIT applies to the contributors’ modifications while the derived Claude Code remains Anthropic’s intellectual property. That sentence deserves a long stare, and gets one below.
Two more names, so you can rule them out. OpenClaw is Peter Steinberger’s self-hosted personal assistant (the project formerly known as Warelay, then Moltbot), a messaging gateway rather than a coding harness; if you searched OpenClaude and meant that, this is the wrong article. Crush is Charm’s terminal coding agent, which began life as a fork in the opencode naming dispute before Charm renamed it and ceded the OpenCode name to the SST project. Four projects, two of them with “claude” or “claw” in the name, none of them affiliated with each other. No wonder search engines scramble them.
Is there a benchmark comparing OpenCode and Claude Code?
No benchmark compares OpenCode and Claude Code directly. As of July 2026, no public leaderboard scores OpenCode or OpenClaude at all. The closest evidence is Terminal-Bench 2.1, which scores the harness and the model as separate columns and includes Claude Code (best entry 83.1%) but neither open-source tool.
That two-column design is what makes Terminal-Bench worth your attention. Most coding leaderboards score a model plus an unnamed scaffold and let the vendor take credit for the sum, a problem this site’s tour of the 2026 agent benchmarks covers at length. Terminal-Bench instead runs each agent-model pair against the same suite of real terminal tasks, so when the same model shows up under two different harnesses, the score gap between those rows is the harness, isolated and quantified.
The leaderboard’s control group is Terminus 2, the reference agent maintained by the Terminal-Bench team itself: a deliberately plain harness that exists so every model has a neutral baseline. Five same-model pairings are readable off the current board, and they are the closest thing to a harness benchmark that exists in public.
| Item | Shipping CLI harness | Terminus 2 (reference agent) |
|---|---|---|
| Claude 5 Fable · Claude Code | 83.1% | 80.4% |
| Opus 4.8 · Claude Code | 78.9% | 74.6% |
| Opus 4.7 · Claude Code | 69.7% | 66.1% |
| GPT-5.5 · Codex CLI | 83.4% | 78.2% |
| Gemini 3.1 Pro · Gemini CLI | 70.7% | 70.3% |
Read the Claude rows together. On Claude 5 Fable, Claude Code scores 83.1% against the reference agent’s 80.4%: 2.7 points from the harness alone. On Opus 4.8 the gap widens to 4.3 points (78.9% vs 74.6%), and on the older Opus 4.7 it is 3.6 (69.7% vs 66.1%). Codex CLI shows the same effect even more strongly, adding 5.2 points to GPT-5.5 over the reference. Gemini CLI, on the other hand, adds roughly nothing to Gemini 3.1 Pro: 70.7% vs 70.3%, well inside the error bars.
Three things follow from that pattern, and each one matters for this comparison.
First, harness quality is real and measurable. A well-built agent loop is worth about as much as a model generation: the 4.3 points Claude Code adds to Opus 4.8 is larger than many model-over-model upgrades. Whatever the difference between a harness and a model sounds like in the abstract, on this board it is worth up to five points.
Second, the premium is not automatic. Gemini CLI proves a first-party harness can add nothing. So “OpenCode is probably about as good” is not a safe default assumption, and neither is “surely a derivative of Claude Code performs like Claude Code.” The spread between the best and worst harness effect on this one board is five points. Where OpenCode and OpenClaude fall inside (or outside) that spread is unknown, because nobody has run them.
Third, the number attached to a tool is really a number attached to a pair. Claude Code’s 83.1% belongs to Claude Code running Claude 5 Fable, a pairing that went dark for several days in June when a US export-control directive suspended the model, then returned with the current leaderboard entry dated June 17. Scores travel with model availability, a lesson the Claude Code vs Codex head-to-head covered when that suspension first scrambled the board.
OpenCode and OpenClaude: not on any benchmark yet
State the negative result plainly. OpenCode has no entry on Terminal-Bench 2.0 or 2.1, no entry on SWE-bench in any variant, and no independently measured score anywhere this site could find as of July 7, 2026. OpenClaude is in the same position, and its repository claims no benchmark results either. Every published comparison of these tools, including the good ones, is unmeasured.
Be careful with what that absence means. It is not evidence that OpenCode is worse. Terminal-Bench submissions are run by whoever cares to submit them, and the entries on the board skew toward vendors with a marketing motive: Anthropic, OpenAI, and Google all benefit from a leaderboard presence, while an MIT-licensed side project has nobody paid to run the eval. Absence of evidence here is mostly absence of a submitter.
But it does mean three practical things. You cannot cite a number for OpenCode, so any claimed performance parity is vibes. You cannot price OpenCode’s token efficiency against Claude Code’s, because tokens-per-task is a harness property and nobody has measured it on a standard suite (the closest data point, Builder.io’s four-task January 2026 test, had Claude Code finish in 9 minutes 9 seconds against OpenCode’s 16 minutes 20 seconds, with OpenCode writing more tests: interesting, and far too small to generalize). And you should discount any ranking that silently mixes measured and unmeasured tools, the same failure mode as grading the same model with two different graders.
The gap is also closing from the research side, which makes the missing entries more conspicuous, not less. A June 2026 arXiv preprint, Claw-SWE-Bench (Zheng et al.), builds a benchmark specifically for evaluating agent harnesses on coding tasks, scoring research harnesses like SWE-Agent, OpenHands, and Mini-SWE-Agent. The field now measures harnesses on purpose. The two open-source tools in this comparison simply have not stepped on the scale.
The comparisons you can make today
Benchmarks are one axis. The other five are verifiable right now, from primary sources, and on several of them the three tools are not even playing the same game.
| Axis | Claude Code | OpenCode | OpenClaude |
|---|---|---|---|
| Benchmark presence | Terminal-Bench 2.1, four entries, best 83.1% | None | None |
| Software cost | Subscription $20 to $200 a month, or API rates | Free, MIT | Free, license caveat below |
| Model access | Anthropic models only | 75+ providers, local models included | 18+ providers, local included |
| Source | Closed | Open, from scratch | Open, Claude Code derivative |
| GitHub adoption | n/a (closed distribution) | 183k stars, 22.7k forks | 29.8k stars, 8.9k forks |
| Legal posture | First party | Clean MIT | “Derived Claude Code remains Anthropic IP” |
Cost: the subscription is a moat, not a price tag
The list prices look simple: Claude Code costs $20 to $200 a month, the other two are free software where you pay per token. The real difference is which rates you are allowed to pay.
Until January 2026, OpenCode users could sign in with a Claude Pro or Max account and run Anthropic models at subscription prices from inside a third-party harness. Then Anthropic closed the door in three moves. On January 9, 2026, server-side checks began rejecting subscription OAuth tokens used outside official surfaces. In February, as The Register reported, Anthropic added an authentication and credential clause to its consumer terms banning exactly that pattern. And on March 19, 2026, OpenCode’s maintainer merged a pull request titled “anthropic legal requests” removing the project’s Anthropic OAuth plugin and Claude system prompt.
The consequence is a hard fork in the economics. Inside Claude Code, a $200 Max plan buys a large multiple of what the same $200 buys at metered API rates, the arithmetic this site ran on Claude Code’s real per-task costs. Inside OpenCode or OpenClaude, Anthropic models are available only at full API prices, which flips the value math: the open harnesses make the most financial sense with the providers where bring-your-own-key is the native billing model, or with local models at zero marginal token cost, the pattern that defines the whole BYOK alternatives field.
So the honest cost comparison is not “paid vs free.” It is: heavy Claude usage is cheapest inside Claude Code, and everything that is not a Claude model is cheapest inside an open harness.
Provider freedom, extensibility, adoption
On model access there is no contest to score. Claude Code runs Anthropic models. OpenCode’s 75+ providers and OpenClaude’s 18+ both cover the OpenAI-compatible universe plus local inference through Ollama; OpenCode’s Models.dev integration is the wider and better-maintained catalog.
Extensibility is closer than the open-vs-closed framing suggests. Claude Code’s primitive set (skills, hooks, subagents, MCP, plugins) is the deepest extension surface of any shipping agent, and it is versioned and documented by a vendor with a support obligation. OpenCode counters with two built-in agents (a full-access build agent and a read-only plan agent), a general subagent, LSP autoloading, and the blunt advantage that all of it is MIT-licensed TypeScript you can read and patch. OpenClaude inherits much of Claude Code’s surface by descent, which sounds like a feature until you remember its upstream is a decompiled moving target it can only chase.
Adoption is where the two open projects separate. OpenCode carries 183k stars, a thousand-odd contributors, and a release cadence that shipped v1.17.14 the day before this article. OpenClaude’s 29.8k stars in a few months is respectable growth, but it is a fraction of the field, and star counts are the weakest of signals anyway; the broader 2026 agent market is better read through usage share than GitHub applause.
Provenance: the OpenClaude question
OpenClaude’s own license framing is the part a lawyer would underline: MIT for the contributors’ modifications, while the derived Claude Code remains Anthropic’s intellectual property. Translated, the project ships code it acknowledges it does not fully own the rights to relicense. Anthropic has already shown, in the OAuth episode above, that it will send legal requests to open-source projects far less entangled with its property than a direct derivative. Nothing public suggests action against OpenClaude as of July 2026. But if your team is choosing infrastructure, “the vendor whose code this is derived from has an active enforcement posture” is a risk line item, and it is one OpenCode simply does not carry.
Verdict: which harness should you pick?
- Pick Claude Code if you run Anthropic models heavily. It is the only one of the three with measured benchmark results (83.1% best entry on Terminal-Bench 2.1, and a proven 2.7 to 4.3 point harness premium over the reference agent), and the only one that can bill through a Claude subscription instead of API rates.
- Pick OpenCode if you want provider freedom, local models, readable source, or a $0 software bill with a clean MIT license. You are accepting unmeasured performance: possibly fine, genuinely unknown.
- Treat OpenClaude as an experiment, not infrastructure. It has no benchmark presence, a tenth of OpenCode’s adoption, and a license posture that concedes its core is Anthropic’s property. Curiosity yes, production no.
- Whatever you pick, run your own eval. A week of your real tasks with tokens logged beats every number in this article, because the published data covers one of the three tools and your workload is on none of the boards.
Frequently asked questions
- Is there a benchmark comparing OpenCode and Claude Code?
- No. As of July 2026, no public benchmark scores OpenCode at all: it has no entry on Terminal-Bench 2.0 or 2.1 or on any SWE-bench variant. Terminal-Bench 2.1 measures Claude Code (best entry 83.1% with Claude 5 Fable), so the two tools have never been scored on the same suite. Published comparisons are hands-on impressions, not benchmarks.
- Is OpenCode better than Claude Code?
- Unknown, and anyone claiming otherwise is guessing. Claude Code has measured results (a 2.7 to 4.3 point premium over the Terminal-Bench reference agent on the same model). OpenCode has no measured results. Hands-on tests like Builder.io in January 2026 found Claude Code faster (9m09s vs 16m20s over four tasks) and OpenCode more thorough, but a four-task test does not generalize.
- What is OpenClaude, and is it the same as OpenClaw?
- No. OpenClaude (Gitlawb/openclaude, 29.8k GitHub stars) is an open-source coding agent CLI derived from the Claude Code codebase and modified to support 18+ providers. OpenClaw is an unrelated self-hosted personal assistant by Peter Steinberger, formerly Warelay and Moltbot, and is not a coding harness.
- Can I use my Claude Pro or Max subscription with OpenCode or OpenClaude?
- No. Anthropic blocked subscription OAuth tokens in third-party tools on January 9, 2026, and wrote the ban into its consumer terms in February 2026. OpenCode removed its Anthropic OAuth plugin in March 2026 after legal requests. Anthropic models in third-party harnesses now bill at API rates only.
- Does the harness really change benchmark scores?
- Yes, measurably. On Terminal-Bench 2.1, holding the model constant, Claude Code scores 2.7 to 4.3 points above the Terminus 2 reference agent and Codex CLI scores 5.2 points above it, while Gemini CLI adds roughly nothing. The harness premium can be worth as much as a model upgrade, which is why unmeasured harnesses are a real unknown.
Sources
Anthropic (2026). Claude pricing. Anthropic (vendor documentation). https://claude.com/pricing
Builder.io (2026, January 12). OpenCode vs Claude Code. Builder.io (hands-on comparison, secondary). https://www.builder.io/blog/opencode-vs-claude-code
Composio (2026, June). OpenCode vs Claude Code (2026): After 100 hours of usage. Composio (hands-on comparison, secondary). https://composio.dev/content/claude-code-vs-open-code
Charm (2026). Crush repository. GitHub. https://github.com/charmbracelet/crush
Gitlawb (2026). OpenClaude repository. GitHub. https://github.com/Gitlawb/openclaude
OpenCode / Anomaly (2026). OpenCode repository and opencode.ai. GitHub and vendor site. https://github.com/sst/opencode
OpenCode / Anomaly (2026, March 19). Pull request #18186: anthropic legal requests. GitHub. https://github.com/anomalyco/opencode/pull/18186
Terminal-Bench (2026). Terminal-Bench 2.1 leaderboard. tbench.ai (independent benchmark). https://www.tbench.ai/leaderboard/terminal-bench/2.1
The Register (2026, February 20). Anthropic clarifies ban on third-party tool access to Claude. The Register (secondary coverage). https://www.theregister.com/2026/02/20/anthropic_clarifies_ban_third_party_claude_access/
Zheng, M. et al. (2026). Claw-SWE-Bench: A Benchmark for Evaluating OpenClaw-style Agent Harnesses on Coding Tasks. arXiv (preprint). https://arxiv.org/abs/2606.12344