April 2026 has been the most consequential month in frontier AI since GPT-4's launch. OpenAI shipped GPT-5.5, Anthropic revealed Claude Mythos Preview to select partners, DeepSeek open-sourced a 1.6-trillion-parameter model, and Grok 4.20 tied GPT-5.4 Pro at the top of the Mensa Norway benchmark. The frontier is crowded, and the gap between first and fifth place is measured in single digits.
Here is a benchmark-driven ranking of the smartest AI models available right now, what each one does best, and where they fall short.
The Top Tier: Intelligence Benchmarks
TrackingAI's April 2026 Mensa Norway benchmark puts Grok 4.20 Expert Mode and GPT-5.4 Pro (Vision) in a dead heat at 145 IQ-equivalent — but raw IQ scores only tell part of the story. When you look at practical benchmarks, the picture gets more nuanced.
Coding and Agentic Work
GPT-5.5 dominates here. Terminal-Bench 2.0, which measures real-world software engineering ability, jumped from 75.1% on GPT-5.4 to 82.7% on GPT-5.5. That is not incremental — it is the single largest generation-over-generation improvement OpenAI has posted on that benchmark. Claude Opus 4.7 sits at 69.4%, and Gemini 3.1 Pro at 68.5%.
For agentic coding tasks — where the model operates autonomously across tools, reads documentation, writes tests, and submits pull requests — GPT-5.5 is the current leader. OpenAI's internal Expert-SWE eval (tasks averaging 20 human hours) shows GPT-5.5 at 73.1%, up from 68.5%.
Claude Mythos Preview reportedly matches or slightly trails GPT-5.5 on Terminal-Bench 2.0, putting it in a statistical tie. But since Mythos is gated and not publicly available, developers cannot verify this independently.
Math and Reasoning
GPT-5.5 Pro leads FrontierMath Tier 4 at 39.6%, followed by GPT-5.5 base at 35.4%. Claude Opus 4.7 manages 22.9%, and Gemini 3.1 Pro drops to 16.7%. The gap is significant: GPT-5.5 Pro nearly doubles Gemini's score on the hardest math benchmark currently available.
DeepSeek V4 Pro has not yet published FrontierMath scores, but based on DeepSeek R1's historical performance on math benchmarks, it is expected to be competitive here.
Web Browsing and Research
BrowseComp tells an interesting story. GPT-5.5 Pro leads at 90.1%, but Gemini 3.1 Pro is close behind at 85.9%, beating GPT-5.5 base (84.4%). Google's model excels at information retrieval and synthesis, which makes sense given their index advantage.
GPT-5.5 base already cleared 84% — a 1.7-point jump over GPT-5.4 — so even without paying Pro pricing, you get strong research capabilities.
Computer Use
OSWorld-Verified measures how well a model can navigate real desktop software. GPT-5.5 leads at 78.7%, barely edging out Claude Opus 4.7 at 78.0%. These two are effectively tied for computer-use tasks, which is notable because Anthropic has invested heavily in this capability. Gemini does not yet publish OSWorld scores.
The Open-Source Challenger: DeepSeek V4
DeepSeek V4, released April 24, changes the economics entirely. Two models — V4-Flash (284B total, 13B active, $0.14/$0.28 per million tokens) and V4-Pro (1.6T total, 49B active, $1.74/$3.48) — deliver frontier-adjacent performance at a fraction of the cost.
V4-Pro's $1.74/$3.48 pricing undercuts GPT-5.5 ($5.00/$30.00) by 3-8x and Claude Sonnet 4.6 ($3.00/$15.00) by 2-4x. For teams running high-volume inference pipelines, this is transformative.
The catch: V4 benchmark scores are still being validated by the community. DeepSeek claims competitive performance with GPT-5.4 on most benchmarks, but independent verification is ongoing. Still, at MIT-licensed open weights, the value proposition is hard to ignore.
The Complete Ranking
| Rank | Model | Best At | Pricing (In/Out per M tokens) | Availability |
|---|---|---|---|---|
| 1 | GPT-5.5 Pro | Deep research, hard math | $15 / $75 | Public |
| 2 | GPT-5.5 | Coding, agentic tasks | $5 / $30 | Public |
| 3 | Claude Mythos Preview | Reasoning (claimed) | Unknown | Gated preview |
| 4 | Grok 4.20 Expert | Raw IQ, general reasoning | $5 / $25 | Public |
| 5 | Claude Opus 4.7 | Computer use, nuance | $15 / $75 | Public |
| 6 | Gemini 3.1 Pro | Web research, multimodal | $2 / $12 | Public |
| 7 | DeepSeek V4 Pro | Cost-efficient reasoning | $1.74 / $3.48 | Open weights |
| 8 | DeepSeek V4 Flash | High-volume tasks | $0.14 / $0.28 | Open weights |
What This Means for Developers
Three practical takeaways from the April 2026 landscape:
For coding, GPT-5.5 is the default. The Terminal-Bench and Expert-SWE improvements are real and measurable. If you are building agentic coding tools, GPT-5.5 should be your starting point.
For cost-sensitive production, DeepSeek V4 is a no-brainer. At $0.14/$0.28 for Flash and $1.74/$3.48 for Pro, you can run 10-100x more inference for the same budget. The open-source ecosystem will fine-tune and optimize these models rapidly.
Watch Claude Mythos closely. Anthropic has not released Mythos publicly, but the benchmark leaks suggest it matches GPT-5.5 on coding and potentially exceeds it on certain reasoning tasks. When (or if) it launches publicly, the rankings shift again.
Looking Ahead
OpenAI is reportedly working on GPT-6, which could land as early as late 2026. Anthropic has Claude Mythos in preview, suggesting a full release within weeks. Google is iterating on Gemini 3.x. And DeepSeek just proved that open-source can compete on price-performance even if it trails slightly on raw intelligence.
The pace is not slowing down. The model you pick today will not be the best model in three months. Build your infrastructure to be model-agnostic — the flexibility to swap providers will matter more than picking any single winner.