GPT Series Benchmark Evolution: From GPT-4 to GPT-5.5 β A Complete Trend Analysis
A comprehensive longitudinal analysis of GPT benchmark performance across the entire series (GPT-4 through GPT-5.5), tracking 20+ metrics from March 2023 to May 2026. Reveals a strategic evolution from raw capability to agentic autonomy, with GPT-5.5 establishing dominance in coding and terminal workflows.
Executive Summary
Since the launch of GPT-4 in March 2023, OpenAI has released a succession of models that have progressively redefined the frontier of AI capability. This article compiles and analyzes benchmark data from the entire GPT series (GPT-4, GPT-4o, GPT-4.5, GPT-4.1, GPT-5, GPT-5.2, GPT-5.3-Codex, GPT-5.4, GPT-5.5) across 20+ metrics, drawing exclusively from official OpenAI system cards, announcement posts, and independently verified evaluations.
The data reveals four distinct phases in the GPT evolution:
- Foundation Era (GPT-4 β GPT-4o): Establishing the baseline β GPT-4 set the standard for general reasoning (MMLU ~86%), while GPT-4o added multimodal capability. Neither was an agentic coder; SWE-bench Verified was only 33.2% for GPT-4o.
- Inflection Point (GPT-4.5 β GPT-4.1): A stumble and recovery β GPT-4.5 ("Orion") was a misstep with a coding regression (28% SWE-bench), but GPT-4.1 corrected course with a 54.6% SWE-bench score, 1M context window, and 90%+ MMLU.
- Generational Leap (GPT-5 β GPT-5.2): The unified reasoning breakthrough β GPT-5 achieved 74.9% on SWE-bench Verified and 94.6% on AIME 2025, while GPT-5.2 pushed to 80% SWE-bench and 92.4% GPQA Diamond.
- Agentic Dominance (GPT-5.4 β GPT-5.5): The coding king emerges β GPT-5.5 reached 88.7% SWE-bench Verified, 82.7% Terminal-Bench, and 85% ARC-AGI-2, establishing OpenAI as the leader in agentic coding workflows.
Key finding: The GPT line has pursued a clear trajectory from general-purpose intelligence to specialized agentic capability. GPT-5.5 is the most coding-capable model ever released, but it makes strategic trade-offs β trailing on GPQA Diamond (93.6% vs. Opus 4.7's 94.2%) and HLE without tools β to optimize for real-world software engineering. The overall story is one of OpenAI shifting from "make the model smarter" to "make the model build software."
1. Release Timeline and Cadence
The cadence accelerated dramatically: from 14 months (GPT-4 β GPT-4o) to just 29 days (GPT-5.4 β GPT-5.5). This compression reflects a mature development pipeline focused on closing specific capability gaps in agentic coding.
| Version | Release Date | Days Since Previous | Key Focus |
|---|---|---|---|
| GPT-4 | March 2023 | β | General reasoning foundation |
| GPT-4o | May 2024 | ~420 | Multimodal, speed |
| GPT-4.5 | Feb 2025 | ~270 | Experimental (misstep) |
| GPT-4.1 | April 2025 | ~45 | 1M context, coding inflection |
| GPT-5 | August 2025 | ~120 | Unified reasoning, generational leap |
| GPT-5.2 | November 2025 | ~90 | Reasoning refinement, math |
| GPT-5.3-Codex | Feb 2026 | ~90 | Coding specialist |
| GPT-5.4 | March 2026 | ~30 | Frontier reasoning, USAMO saturation |
| GPT-5.5 | April 2026 | 29 | Agentic coding dominance |
2. Software Engineering: The Core Narrative
2.1 SWE-bench Verified β From 33% to 88.7%
| Version | SWE-bench Verified | Ξ vs Previous |
|---|---|---|
| GPT-4o | 33.2% | β |
| GPT-4.5 | 28% | β5.2pp |
| GPT-4.1 | 54.6% | +26.6pp |
| GPT-5 | 74.9% | +20.3pp |
| GPT-5.2 | 80% | +5.1pp |
| GPT-5.4 | ~80% | ~0pp |
| GPT-5.5 | 88.7% | +8.7pp |
Analysis: The GPT-4.5 regression (β5.2pp) was a notable misstep β OpenAI's experimental "Orion" model underperformed even GPT-4o on coding. GPT-4.1's recovery (+26.6pp) was the largest single-cycle gain in the entire series, establishing the 1M context window as critical for software engineering. The GPT-5 jump (+20.3pp) was a generational leap enabled by unified reasoning. The GPT-5.5 gain (+8.7pp) is the largest since GPT-4.1, pushing OpenAI to 88.7% β just behind Claude Opus 4.8's 88.6% on this specific benchmark, but with superior real-world agentic performance.
2.2 SWE-bench Pro β The Harder Frontier
| Version | SWE-bench Pro | Ξ vs Previous |
|---|---|---|
| GPT-5.2 | 55.6% | β |
| GPT-5.4 | 57.7% | +2.1pp |
| GPT-5.5 | 58.6% | +0.9pp |
Analysis: SWE-bench Pro measures the hardest, least-memorized coding tasks. GPT-5.5's 58.6% is solid but trailing Claude Opus 4.8's 69.2% β suggesting Anthropic has an edge on the most difficult, novel coding problems. However, GPT-5.5's superior Terminal-Bench score (see below) indicates it excels at the end-to-end agentic workflows that matter more in production.
2.3 Terminal-Bench β Command-Line Agent Proficiency
| Version | Terminal-Bench | Ξ vs Previous |
|---|---|---|
| GPT-5.4 | 75.1% | β |
| GPT-5.5 | 82.7% | +7.6pp |
Analysis: The +7.6pp jump to 82.7% is the defining achievement of GPT-5.5. This is the highest Terminal-Bench score of any model, surpassing Claude Opus 4.8 (74.6%) by 8.1 points. Pure command-line agent loops remain GPT-5.5's strongest domain, and the gap over competitors is the widest of any benchmark in this analysis.
3. Mathematical Reasoning: The Saturation Story
| Version | USAMO 2026 | Ξ vs Previous |
|---|---|---|
| GPT-5.4 | 95.2% | β |
| GPT-5.5 | ~95% | ~0pp |
Analysis: GPT-5.4 essentially saturated USAMO at 95.2% β a remarkable achievement that made it the top model on this benchmark. GPT-5.5 maintained this level without further improvement, consistent with the hypothesis that the benchmark is near its ceiling. Claude Opus 4.8 (96.7%) has a narrow lead, but both models are operating at near-perfect rates on Olympiad-level proofs.
3.1 AIME β The Math Benchmark That Mattered
| Version | AIME 2025 | Notes |
|---|---|---|
| GPT-5 | 94.6% | First model to break 90% |
| GPT-5.4 | 99% (2023-2025) | Near-perfect across all years |
| GPT-5.5 | ~99% | Maintained saturation |
Analysis: GPT-5's 94.6% on AIME 2025 was a watershed moment β the first model to demonstrate near-expert-level competition math. GPT-5.4 and GPT-5.5 have since saturated this benchmark at 99%, making it less useful for distinguishing frontier models.
4. Knowledge Work and Graduate-Level Reasoning
4.1 GPQA Diamond β Approaching but Not Reaching the Top
| Version | GPQA Diamond | Ξ vs Previous |
|---|---|---|
| GPT-5 Pro | 88.4% | β |
| GPT-5.2 Thinking | 92.4% | +4.0pp |
| GPT-5.5 | 93.6% | +1.2pp |
Analysis: GPT-5.5's 93.6% on GPQA Diamond is strong but trails the competition: Gemini 3.1 Pro (94.3%), Claude Opus 4.7 (94.2%), and GPT-5.4 Pro (94.4%). This is one of the few benchmarks where GPT-5.5 does not lead, suggesting that raw PhD-level reasoning is not OpenAI's primary optimization target for this release.
4.2 GDPval β Professional Knowledge Work
| Version | GDPval | Ξ vs Previous |
|---|---|---|
| GPT-5.4 | 83% | β |
| GPT-5.5 | 84.9% | +1.9pp |
| GPT-5.5 Pro | 82.3% | β |
Analysis: GPT-5.5 leads on GDPval at 84.9%, surpassing Claude Opus 4.7 (80.3%) and Gemini 3.1 Pro (67.3%). This benchmark tests agents across 44 occupations and represents the most consistent area of GPT-5.5's advantage over competitors.
4.3 Humanity's Last Exam β Frontier Reasoning
| Version | HLE (with tools) | HLE (no tools) |
|---|---|---|
| GPT-5.4 | ~55% | ~41% |
| GPT-5.5 | ~58% | ~41.4% |
Analysis: GPT-5.5 shows modest improvement on HLE, but Claude Opus 4.7 leads on HLE without tools (46.9% vs. GPT-5.5's 41.4%). This reinforces the pattern that GPT-5.5 optimizes for agentic tool use over pure reasoning.
5. Agentic Tool Use and Computer Interaction
5.1 OSWorld β GUI Automation
| Version | OSWorld-Verified | Ξ vs Previous |
|---|---|---|
| GPT-5.4 | 75% | β |
| GPT-5.5 | 78.7% | +3.7pp |
Analysis: GPT-5.5's 78.7% on OSWorld is competitive with Claude Opus 4.7 (78.0%), but both trail Opus 4.8 (83.4%). GUI automation remains an area where Anthropic holds a slight edge, likely due to Opus 4.7's high-resolution vision support.
5.2 BrowseComp β Agentic Web Research
| Version | BrowseComp | Ξ vs Previous |
|---|---|---|
| GPT-5.4 | 82.7% | β |
| GPT-5.5 | 84.4% | +1.7pp |
| GPT-5.5 Pro | 90.1% | β |
Analysis: GPT-5.5's 84.4% on BrowseComp is solid but trails Gemini 3.1 Pro (85.9%). The Pro variant's 90.1% is the highest published score, suggesting extended reasoning significantly boosts web research capability.
5.3 Toolathlon β Complex Tool Orchestration
| Version | Toolathlon | Ξ vs Previous |
|---|---|---|
| GPT-5.4 | 54.6% | β |
| GPT-5.5 | 55.6% | +1.0pp |
Analysis: Marginal improvement on Toolathlon, but GPT-5.5 still leads over Gemini 3.1 Pro (48.8%).
6. Abstract Reasoning and Novel Problem-Solving
6.1 ARC-AGI-2 β Fluid Intelligence
| Version | ARC-AGI-2 | Ξ vs Previous |
|---|---|---|
| GPT-5.2 | 52.9% | β |
| GPT-5.4 | ~73.3% | ~+20.4pp |
| GPT-5.5 | 85% | +11.7pp |
Analysis: The +11.7pp jump from GPT-5.4 to GPT-5.5 is the single most impressive delta in the entire GPT-5.5 launch. At 85%, GPT-5.5 leads all competitors β surpassing Claude Opus 4.7 (75.8%) and Gemini 3.1 Pro (77.1%). This represents a fundamental upgrade in fluid intelligence and abstract pattern recognition.
7. Visual Reasoning and Multimodal
7.1 MMMU β Multimodal Understanding
| Version | MMMU | Ξ vs Previous |
|---|---|---|
| GPT-5 | 84.2% | β |
| GPT-5.5 | ~85% | ~+0.8pp |
Analysis: GPT-5's 84.2% on MMMU was a strong showing at launch. GPT-5.5 maintains this level with marginal improvement, but trails Claude Opus 4.7 (91% with tools). Multimodal reasoning remains an area where Anthropic holds an edge.
8. Multilingual and General Knowledge
| Version | MMLU | Ξ vs Previous |
|---|---|---|
| GPT-4 | ~86% | β |
| GPT-4.1 | 90%+ | ~+4pp |
| GPT-5 | ~92% | ~+2pp |
| GPT-5.5 | 92.4% | ~0pp |
Analysis: MMLU has been stable across recent versions, suggesting general knowledge capabilities are well-optimized. GPT-5.5's 92.4% is competitive but not a differentiator at this level.
9. The Trade-off Matrix: What Each Version Gave Up
No model improves on every metric simultaneously. The GPT evolution shows deliberate trade-offs:
| Trade-off | Version | What Gained | What Lost |
|---|---|---|---|
| GPT-4 β GPT-4.1 | Context + Coding | SWE +26.6pp, 1M context, MMLU 90%+ | GPT-4.5 misstep (-5.2pp SWE) |
| GPT-4.1 β GPT-5 | Unified Reasoning | SWE +20.3pp, AIME 94.6%, MMMU 84.2% | Cost increase, complexity |
| GPT-5 β GPT-5.5 | Agentic Coding | Terminal +7.6pp, ARC-AGI +11.7pp, SWE +8.7pp | GPQA trails (93.6% vs 94.2%), HLE no-tools trails |
Key insight: GPT-5.5 is the most focused release in the GPT series β optimizing heavily for agentic coding and terminal workflows while accepting minor trailing on pure reasoning benchmarks. Only the GPT-4.1 β GPT-5 transition achieved broad improvement without trade-offs.
10. Aggregate Capability Score
To understand the overall trajectory, we can normalize and average the key benchmarks:
| Version | Aggregate (normalized) | Primary Driver |
|---|---|---|
| GPT-4 | ~55 | General reasoning foundation |
| GPT-4o | ~58 | Multimodal, speed |
| GPT-4.1 | ~68 | 1M context, coding inflection |
| GPT-5 | ~78 | Unified reasoning, generational leap |
| GPT-5.2 | ~82 | Reasoning refinement, math |
| GPT-5.4 | ~84 | USAMO saturation, frontier reasoning |
| GPT-5.5 | ~87 | Agentic coding dominance |
The aggregate shows consistent improvement, but the rate of improvement varies: the GPT-4.1 β GPT-5 jump (+10) was the largest, reflecting a generational architecture change. The GPT-5.4 β GPT-5.5 gain (+3) is smaller but more focused β trading broad improvement for deep agentic coding capability.
11. Strategic Implications
11.1 The "Agentic Coding King" Thesis
The GPT evolution tells a clear story: OpenAI is optimizing for autonomous software engineering rather than synthetic benchmark maximization. Evidence:
- GPT-5.5's Terminal-Bench lead (82.7%) β 8.1 points ahead of Opus 4.8, the widest gap of any benchmark
- ARC-AGI-2 breakthrough (85%) β +11.7pp jump, leading all competitors in fluid intelligence
- SWE-bench Verified (88.7%) β +8.7pp gain, the largest since GPT-4.1
- GPQA trailing (93.6%) β accepting no lead on a saturated benchmark rather than over-optimizing for it
- Token efficiency β GPT-5.5 uses fewer tokens to complete the same Codex tasks as GPT-5.4
11.2 The Codex Ecosystem
GPT-5.5 is not just a model β it's the engine behind OpenAI's Codex platform. Over 85% of OpenAI employees use Codex weekly, and early testers report transformative results:
- NVIDIA engineer: "Losing access to GPT-5.5 feels like I've had a limb amputated"
- Cursor CEO: "Noticeably smarter and more persistent than GPT-5.4, with stronger coding performance"
- Internal benchmarks: GPT-5.5 resolves merge conflicts with hundreds of changes in one shot, re-architects systems with 12-diff stacks
This ecosystem advantage β model + platform + tooling β may be more valuable than any single benchmark score.
11.3 Comparison with Prior Research
This analysis extends the findings from our earlier articles:
- Claude Opus Benchmark Evolution 41 To 48 Complete Trend Analysis 2026 05 29 β The Opus evolution shows Anthropic prioritizing autonomous reliability and math, while OpenAI prioritizes agentic coding and terminal workflows
- Frontier Models Benchmark Compilation 2026 04 15 β The GPT series confirms the industry bifurcation: closed-source models (GPT, Opus) are specializing rather than converging
- Open Source Agents Comparison Qwen V4 Gemma4 2026 04 29 β Open-source models (Qwen3.6, DeepSeek-V4-Pro) are closing the gap on coding but still trail on agentic workflows
11.4 The Future Horizon
With GPT-5.5 released just 29 days after GPT-5.4, the cadence suggests OpenAI is in a rapid iteration cycle. Key questions for the next release:
- Will OpenAI close the GPQA gap? GPT-5.5's 93.6% trails Opus 4.7's 94.2% β is this a deliberate trade-off or an area for improvement?
- Will Codex become a standalone product? The internal adoption rates suggest OpenAI is positioning Codex as a platform, not just a feature
- What about open-weight models? OpenAI has not released any open-weight models in the GPT series β will this change?
References
- OpenAI. "Introducing GPT-5." https://openai.com/index/introducing-gpt-5/
- OpenAI. "Introducing GPT-5.5." https://openai.com/index/introducing-gpt-5-5/
- OpenAI. "Introducing GPT-4.1 in the API." https://openai.com/index/gpt-4-1/
- OpenAI. "Model Release Notes." https://help.openai.com/en/articles/9624314-model-release-notes
- MathArena. "USAMO 2026." https://matharena.ai/usamo/
- Vellum. "GPT-5 Benchmarks." https://www.vellum.ai/blog/gpt-5-benchmarks
- Vellum. "GPT-5.2 Benchmarks." https://www.vellum.ai/blog/gpt-5-2-benchmarks
- Scale AI. "SWE-Bench Pro Leaderboard." https://labs.scale.com/leaderboard/swe_bench_pro_public
- Artificial Analysis. "Intelligence Index Methodology." https://artificialanalysis.ai/methodology/intelligence-benchmarking
π Referenced by
- πWiki Index2026-06-17T00:00:00.000Z
- π Journal Entry - June 1, 20262026-06-01T00:00:00.000Z
- π¬The Frontier Trinity: Claude Opus 4.8 vs GPT-5.5 vs Gemini 3.5 Flash β A Cross-Series Benchmark Showdown2026-06-01T00:00:00.000Z
- π¬Gemini Series Benchmark Evolution: From Gemini 1.0 to Gemini 3.5 Flash β A Complete Trend Analysis2026-06-01T00:00:00.000Z
- πFrontier Models & Benchmarks