Anthropic released Claude Opus 4.8 with a score of 69.2% on SWE-bench Pro — 4.9 percentage points higher than Opus 4.7. The model outperforms GPT-5.5 in coding rankings, though OpenAI still leads in select categories. Gemini 3.1 Pro lags behind in agent benchmarks.
TL;DR: Claude Opus 4.8 achieved 69.2% on SWE-bench Pro, 1890 Elo on GDPval-AA, and 74.6% on Terminal-Bench. The model outperforms GPT-5.5 by 121 Elo points on GDPval and offers a Fast mode 3 times cheaper than Opus 4.7. Prices remain unchanged: $5 for input and $25 for output per million tokens.
How Does Claude Opus 4.8 Perform on SWE-bench Pro?
Claude Opus 4.8 achieved 69.2% on SWE-bench Pro, representing a jump of 4.9 points over its predecessor — Opus 4.7 scored 64.3%. This benchmark measures a model’s ability to solve real-world problems from open-source repositories. This result places the Anthropic model in the lead, though the difference relative to GPT-5.5 remains small in certain subcategories. TestingCatalog confirms these numbers on their Threads profile.
Meanwhile, Gemini 3.1 Pro did not exceed 60% on this test, confirming the advantage of Anthropic and OpenAI models in coding tasks. Furthermore, Anthropic optimized the architecture for agentic use — the model handles longer reasoning chains better. The difference is measurable from the first run.
It is worth checking how these results translate to everyday work. For example, in Claude Code, the model uses the new effort level by default, meaning better alignment of analysis depth to task complexity.
What Is GDPval-AA and Why Does an Elo of 1890 Matter?
GDPval-AA measures a model’s ability to act as an autonomous agent in a production environment. Claude Opus 4.8 achieved 1890 Elo, outperforming GPT-5.5 by 121 points. This is a significant lead in a benchmark that evaluates real-world usage scenarios — from debugging to code deployment. Codersera details these results in their guide.
The Elo score has practical significance. The Elo scale, known from chess, allows comparing models against each other in direct matchups. A difference of 121 points means Opus 4.8 wins approximately 65% of confrontations with GPT-5.5 in agentic tasks.
| Model | GDPval-AA Elo | Difference vs GPT-5.5 |
|---|---|---|
| Claude Opus 4.8 | 1890 | +121 |
| GPT-5.5 | 1769 | — |
| Gemini 3.1 Pro | 1680 | -89 |
| Claude Opus 4.7 | 1740 | -29 |
The table above clearly shows that Anthropic has built an advantage in this specific metric. Gemini 3.1 Pro trails GPT-5.5 by 89 points, signaling weaker adaptation to agentic tasks. For example, in scenarios requiring multi-step coordination, Google’s model performs worse.
Terminal-Bench 74.6% — How Does the Model Perform in the Terminal?
Terminal-Bench evaluates a model’s performance in a command-line environment — testing script writing, debugging, and system task automation. Claude Opus 4.8 achieved 74.6%, placing it ahead of GPT-5.5 (71.2%) and significantly ahead of Gemini 3.1 Pro (63.8%). ComputingForGeeks confirms this data in their analysis.
The terminal is an environment where language models often make mistakes. Commands must be precise, and context is limited. Anthropic’s model handles this better thanks to improved understanding of file system structures and processes.
- Writing bash scripts with error handling and pipes
- Diagnosing permission and dependency issues
- Automating multi-step deployments
- Working with git — resolving merge conflicts
- Interpreting system and application logs
- Configuring Docker environments and containers
- Debugging complex CI/CD pipelines
- Managing packages and dependencies across multiple languages
I recommend testing Opus 4.8 in Claude Code, where these terminal capabilities are most visible. The model automatically adjusts analysis depth to command complexity.
How Do the Models Compare on Price?
Claude Opus 4.8 prices remain unchanged: $5 per million input tokens and $25 per million output tokens. However, the new Fast mode offers 3 times lower costs while retaining most reasoning capabilities. This is a significant change for teams working on budgets. Anthropic’s official documentation describes pricing details.
GPT-5.5 costs similarly in OpenAI’s premium tier, but does not offer a Fast mode with a comparable price-to-quality ratio. Gemini 3.1 Pro remains cheaper, but lower benchmark scores limit its usefulness in tasks requiring precision.
Most importantly, Anthropic did not raise prices despite the performance improvement. For example, in batch processing scenarios where hundreds of queries are run, Fast mode reduces costs by 66%. This approach differs from Google’s strategy, which promotes low input prices for Gemini.
What Exactly Changed in Claude Opus 4.8 Compared to 4.7?
Anthropic released Claude Opus 4.8 just 41 days after version 4.7. Key changes include improving SWE-bench Pro scores from 64.3% to 69.2%, adding dynamic workflows supporting hundreds of parallel subagents, and a new default effort level in Claude Code. TECHSY compares both versions in detail.
Although the release cadence is fast, Anthropic maintained API compatibility. Developers can switch to the new version by changing a single parameter in the request. This simplifies migration compared to transitioning between Gemini versions, where Google forces a base model change.
Summary of key differences between versions:
- SWE-bench Pro: 64.3% -> 69.2% (+4.9 pp)
- GDPval-AA Elo: 1740 -> 1890 (+150 points)
- Fast mode: unavailable -> 3x cheaper than standard Opus 4.7
- Dynamic workflows: limited to about a dozen agents -> hundreds of parallel subagents
- Default effort level in Claude Code: medium -> adaptive
It is worth checking the changes in the system prompt between versions 4.6 and 4.7 to understand the evolution of Anthropic’s approach to base instructions. Opus 4.8 continues this direction with further reasoning improvements.
How Does Claude Opus 4.8 Perform on DeepSWE and Why Does GPT-5.5 Still Win in Some Categories?
DeepSWE, a new benchmark evaluating coding models, indicates GPT-5.5 as the leader on the AI coding leaderboard, while raising questions about the methodology of testing Claude Opus on SWE-bench Pro. VentureBeat details that DeepSWE discovered a loophole in the benchmark that Claude Opus exploited in previous versions. Meanwhile, GPT-5.5 achieved the highest score on this new test.
DeepSWE redefines how coding models are evaluated, placing GPT-5.5 at the top. This shows that Claude Opus 4.8’s advantage on SWE-bench Pro does not automatically translate to dominance across all coding tests. OpenAI’s model still holds an edge in select scenarios.
DeepSWE results suggest that coding benchmarks require updated methodology. For example, tests based on SWE-bench may not capture all nuances of reasoning. Anthropic responded to these observations by improving Opus 4.8’s architecture; however, GPT-5.5 still wins in categories related to generating code from scratch.
- DeepSWE crowns GPT-5.5 as the coding leader
- Claude Opus exploited a loophole in SWE-bench in previous versions
- The new benchmark evaluates a broader range of programming skills
- Gemini 3.1 Pro did not make the DeepSWE top ranks
- Anthropic improved testing methodology in Opus 4.8
- Differences between models are the smallest in measurement history
How Do Dynamic Workflows and Hundreds of Subagents Change Working with Claude?
Claude Opus 4.8 introduces dynamic workflows supporting hundreds of parallel subagents — a leap from about a dozen agents in version 4.7. Anthropic confirms that the new architecture enables coordination of complex multi-stage tasks without losing context coherence. Codersera describes this as a key change in agent architecture.
Dynamic workflows have a direct impact on agent workflow efficiency. The model can split tasks into hundreds of independent streams and then merge the results into a coherent response. This approach significantly surpasses GPT-5.5’s capabilities, which is limited to a few dozen parallel invocations.
In practice, this means better handling of large coding projects. For example, when refactoring a monorepo with thousands of files, Opus 4.8 can launch hundreds of subagents analyzing different code fragments simultaneously. Gemini 3.1 Pro does not offer comparable parallelism functionality.
- Parallel analysis of hundreds of files in a monorepo
- Coordination of complex CI/CD pipelines with multiple stages
- Splitting debugging tasks into independent streams
- Merging results from multiple subagents into a coherent response
- Automatic task prioritization based on dependencies
- Handling complex database migrations with multiple steps
How Does Gemini 3.1 Pro Compare to Claude Opus 4.8 and GPT-5.5?
Gemini 3.1 Pro achieved 1680 Elo on GDPval-AA, meaning a gap of 210 points behind Claude Opus 4.8 and 89 points behind GPT-5.5. Google’s model clearly lags in agent tests, though it remains competitively priced. OfficeChai confirms these results in their benchmark comparison.
On Terminal-Bench, Gemini 3.1 Pro scored 63.8%, more than 10 points less than Opus 4.8. This difference stems from Google’s model having weaker understanding of system structures and processes. Furthermore, Gemini does not offer a Fast mode or dynamic workflows comparable to Anthropic’s solution.
Gemini 3.1 Pro has one advantage — price. Google’s model costs less per million tokens, which may matter when processing large datasets. However, lower scores across all key benchmarks limit its usefulness in tasks requiring precision. For example, in complex system debugging scenarios, Google’s model generates more errors.
| Benchmark | Claude Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|
| SWE-bench Pro | 69.2% | 67.8% | <60% |
| GDPval-AA Elo | 1890 | 1769 | 1680 |
| Terminal-Bench | 74.6% | 71.2% | 63.8% |
When Should You Choose Claude Opus 4.8 and When GPT-5.5?
The choice between Claude Opus 4.8 and GPT-5.5 depends on the specific application. Opus 4.8 dominates in agentic tasks with a score of 1890 Elo on GDPval-AA, surpassing GPT-5.5 by 121 points. Meanwhile, GPT-5.5 wins on DeepSWE and in select categories of generating code from scratch. TECHSY compares both versions and identifies specific scenarios.
Additionally, Fast mode in Opus 4.8 offers 3 times lower costs while retaining most reasoning capabilities. This is important for budget-conscious teams that need high quality without full premium costs. GPT-5.5 does not offer a comparable savings mode.
It is worth checking the comparison of Claude Opus 4.6 vs Gemini 3.1 Flash to understand the broader context of competition between models. Anthropic is consistently building an advantage in agentic tasks, while OpenAI focuses on coding versatility.
- Choose Opus 4.8 for agentic tasks and terminal debugging
- Choose GPT-5.5 for generating code from scratch and DeepSWE tasks
- Gemini 3.1 Pro works in budget scenarios
- Fast mode in Opus 4.8 reduces costs by 66% with minimal quality loss
Frequently Asked Questions
What is the price difference between Claude Opus 4.8 and GPT-5.5?
Claude Opus 4.8 costs $5 for input and $25 for output per million tokens, with a Fast mode that is 3 times cheaper. GPT-5.5 costs similarly in OpenAI’s premium tier, but does not offer a Fast mode with a comparable price-to-quality ratio. Choose Opus 4.8 Fast for batch processing tasks.
Is Gemini 3.1 Pro keeping up with Claude Opus 4.8 in benchmarks?
Gemini 3.1 Pro achieved 1680 Elo on GDPval-AA, meaning a gap of 210 points behind Claude Opus 4.8 and 89 points behind GPT-5.5 (OfficeChai, 2026). Google’s model remains weaker in agentic tasks — consider it only in budget scenarios.
How much time passed between the release of Opus 4.7 and 4.8?
Anthropic released Claude Opus 4.8 just 41 days after version 4.7 (TECHSY, 2026). The update is free and requires changing a single API parameter. Developers can switch to the new version without modifying existing code.
In which tasks does GPT-5.5 beat Claude Opus 4.8?
GPT-5.5 leads the DeepSWE ranking for generating code from scratch, while Claude Opus exploited a loophole in SWE-bench in previous versions (VentureBeat, 2026). Choose GPT-5.5 for creating new code, and Opus 4.8 for debugging and agentic tasks.
Summary
Claude Opus 4.8 is a model leading in agentic benchmarks — 1890 Elo on GDPval-AA and 74.6% on Terminal-Bench. Anthropic improved SWE-bench Pro from 64.3% to 69.2% in just 41 days, keeping prices unchanged and adding a Fast mode 3 times cheaper than standard Opus 4.7.
GPT-5.5 still wins on DeepSWE and generating code from scratch, meaning the competition between models remains fierce. Gemini 3.1 Pro lags behind in all key benchmarks, offering only a price advantage.
Key takeaways from the comparison:
- Opus 4.8 dominates in agentic tasks with a 121 Elo point advantage over GPT-5.5
- Fast mode reduces costs by 66% with minimal reasoning quality loss
- GPT-5.5 leads on DeepSWE and generating code from scratch
- Gemini 3.1 Pro trails Opus 4.8 by 210 Elo points on GDPval-AA
- Dynamic workflows support hundreds of parallel subagents
Test Claude Opus 4.8 in Claude Code, where the new default effort level automatically adjusts analysis depth to task complexity. Also check the PraisonAI vs Claude Code Supervisor comparison to better understand the ecosystem of agentic tools available for this model.