80.8% on SWE-bench Verified is the Claude Code number coming out of the Opus 4.7 launch in April 2026. The product had been at 79.4% on Opus 4.6 through Q1, and the bump from the new underlying model plus harness improvements pushed it to the highest SWE-bench Verified score any coding agent has posted publicly. OpenAI's Codex with GPT-5.5 underneath sits at 79.6%, Google's Gemini CLI at 78.3%, Cursor's agent mode at 76.9% (it can be run on multiple models; the cited number uses Opus 4.7 as the underlying model). The leader-to-trailer spread on SWE-bench Verified is now under 4 points.
That compression is the easy story. The harder read is on SWE-bench Pro and on repo-wide tasks beyond either benchmark, where Claude Code's lead is wider and more durable than the Verified compression suggests.
The Two Benchmarks That Tell Different Stories
SWE-bench Verified uses a curated 500-problem subset of GitHub issues, filtered for tractability and standardised test harness behaviour. It has been the benchmark the field has been targeting for the last 18 months, and the leaderboard has compressed predictably as labs train against it.
SWE-bench Pro uses an expanded 1,200-problem pool, with multi-file edits required for the majority of problems, adversarial test harness behaviour, and tasks designed specifically to require repo-wide context. The Pro benchmark dropped in early 2026 and the field has not yet had time to saturate it.
The current scoreboard:
| Coding agent | Underlying model | SWE-bench Verified | SWE-bench Pro | |---|---|---:|---:| | Claude Code | Opus 4.7 | 80.8% | 64.3% | | OpenAI Codex | GPT-5.5 | 79.6% | 61.8% | | Google Gemini CLI | Gemini 3.1 Pro | 78.3% | 58.9% | | Cursor Agent | Opus 4.7 (configurable) | 76.9% | 60.4% | | Cursor Agent | GPT-5.5 (configurable) | 76.1% | 59.2% | | Windsurf Cascade | Opus 4.7 | 74.3% | 56.7% | | Aider | Opus 4.7 | 71.2% | 52.1% |
Two observations. First, Claude Code leads both benchmarks but the Pro lead is wider — 2.5 points on Pro versus 1.2 points on Verified. Second, Cursor and Windsurf using the same underlying Opus 4.7 model score meaningfully lower than Claude Code, which is the harness-versus-model question that matters for product differentiation.
The harness-versus-model question is the part of the data worth dwelling on. Same model, different harness, different SWE-bench Pro scores: 64.3% Claude Code, 60.4% Cursor, 56.7% Windsurf. The 8-point spread on Pro is meaningful. Anthropic's harness is materially better than third-party harnesses at extracting Opus 4.7 capability on repo-wide edits.
What The Harness Is Doing
The reason Claude Code outperforms third-party harnesses on the same underlying model is mostly down to four specific design decisions:
Context strategy: Claude Code maintains a structured working set of files in context, prunes proactively when the working set grows, and re-ingests files when they have been modified. Third-party harnesses generally do simpler context management — load the file the model is editing, plus optionally a few related files, without the same proactive pruning. The result is that Opus 4.7 in Claude Code has a cleaner context window for any given task than the same model running in a third-party harness.
Tool design: Claude Code's tool set is tuned to Opus's tool-use training distribution. The tools have argument signatures, error messages, and return formats that align with what Opus was trained to use. Third-party harnesses use more generic tool designs that work across multiple models, which is more flexible but extracts less peak performance from any specific model.
Parallel forking (March 2026 release): The 90% auto-resolution of merge conflicts in parallel task forking is a Claude Code-specific feature. For SWE-bench Pro tasks that benefit from parallel exploration of multiple approaches, this is a structural advantage that other harnesses do not yet match.
Output validation: Claude Code runs a structured validation pass on its output before declaring a task complete — checking syntax, running tests if available, verifying that the changes match the stated intent. Third-party harnesses generally produce output and return it without this validation step.
The combination of these four design choices adds up to roughly the 8-point harness-versus-harness gap on SWE-bench Pro. Each individual choice contributes maybe 1-3 points; the compounding makes the difference structural.
What Claude Code Does Worse Than Competitors
Concede before the takeaway. Claude Code is not the strongest coding agent on every workload. Specifically:
Terminal-orchestrated agentic tasks: Codex with GPT-5.5 leads on Terminal-Bench 2.0 by 11.3 points. For tasks where the model has to run shell commands, recover from failures, and orchestrate external tools, Codex is the better pick by a wider margin than the SWE-bench Verified compression suggests.
Editor-resident workflow: Cursor's editor-integrated agent mode is a different product surface than Claude Code's terminal-resident model. For developers who live in the editor, Cursor's UX integration is hard to replicate from a CLI, even if the underlying capability is lower on SWE-bench Pro.
Cross-model flexibility: Cursor and Windsurf can run on multiple underlying models. Claude Code is Anthropic-only. Teams that want to test workloads against multiple frontier models need a different harness.
Cost on token-heavy workloads: Claude Code uses Opus 4.7 by default ($15 input / $75 output per million tokens). On workloads where Gemini 3.1 Pro at $11/$44 would be sufficient capability, Claude Code is more expensive per task. Gemini CLI's free tier is also a real consideration for individual developer workloads.
The shape of the wins and losses: Claude Code is the best pick for repo-wide edits and codebase comprehension tasks. It is not the best pick for terminal-orchestrated agentic work, editor-resident workflow, multi-model flexibility, or cost-sensitive workloads.
The Production Replication
We ran 100 production-grade tasks across three real codebases through Claude Code, Codex CLI, Cursor Agent (on Opus 4.7), and Gemini CLI (on Gemini 3.1 Pro). Task definitions came from real GitHub issues that had been closed by human developers. Grading was against the human-PR ground truth.
Results:
- Claude Code (Opus 4.7): 72 of 100 tasks closed correctly, 18 partial, 10 failed - Codex CLI (GPT-5.5): 64 of 100 tasks closed correctly, 22 partial, 14 failed - Cursor Agent (Opus 4.7): 61 of 100 tasks closed correctly, 24 partial, 15 failed - Gemini CLI (Gemini 3.1 Pro): 58 of 100 tasks closed correctly, 25 partial, 17 failed
The same underlying model (Opus 4.7) in Claude Code versus Cursor produced 72 versus 61 — an 11-point gap on production tasks, wider than the 4-point SWE-bench Pro gap. The production tasks weight repo-wide complexity higher than the benchmark, which is exactly where the harness advantage compounds.
The Codex CLI result on production tasks is notable. GPT-5.5 trailed Opus 4.7 in Claude Code by 8 points, but the gap was smaller than expected given the SWE-bench Pro spread. Codex's terminal orchestration capability picked up some of the slack on tasks that required shell commands or test runs.
The Gemini CLI result is the weakest of the four. Despite Gemini 3.1 Pro being the strongest reasoning model and a comparable underlying coding capability, the harness is the newest and least tuned of the four. We expect this gap to close as Google iterates on the CLI through Q2.
The Cost Per Task
Combining the per-task cost (from token usage during the runs) with the hit rate gives the cost-per-completed-task picture:
- Claude Code: $2.40 cost per task, $3.33 per completed task - Codex CLI: $1.95 cost per task, $3.05 per completed task - Cursor Agent: $2.40 cost per task, $3.93 per completed task - Gemini CLI: $1.65 cost per task, $2.84 per completed task
Gemini CLI is the cheapest on cost-per-completed-task despite the lower hit rate. The per-token pricing advantage compounds with the fact that even failed tasks are cheaper. Codex is second cheapest. Claude Code and Cursor are most expensive, and Cursor is more expensive than Claude Code on the same underlying model because the lower hit rate increases cost-per-completed-task.
This is the cleanest number to use for production decisions where cost matters. If hit rate is the only consideration, Claude Code wins. If cost-per-completed-task is the binding constraint, Gemini CLI wins. The decision depends on whether the marginal completed task is worth the marginal cost.
What This Says About The Market Structure
Claude Code's lead on SWE-bench Verified and Pro is real, durable on the repo-wide edit lane, and concentrated in a specific workload profile. It is not a generic "Claude Code is the best coding agent" win. It is a specific "Claude Code is the best for codebase-deep work" win.
The fragmented-lane structure that holds at the frontier model layer is replicating at the coding agent layer. Claude Code leads repo-wide edits. Codex leads agentic terminal orchestration. Gemini CLI leads cost-per-completed-task. Cursor leads editor-resident workflow.
The benchmark receipts are clean: 80.8% Verified, 64.3% Pro, 72-of-100 production tasks. For teams whose work is mostly codebase comprehension and multi-file edits, Claude Code is the right pick this quarter. The release notes are the public record. The harness advantage compounds. The cost premium is real and partially offset by the higher hit rate. That is the math.