April 2026's LLM War: Benchmarks, Tokens, Dollars

One week separated two of the most consequential model launches in recent memory. Anthropic released Claude Opus 4.7 on April 16, 2026, pushing SWE-bench Pro from 53.4% to 64.3% and adding high-resolution vision at the same $5 per million input tokens as its predecessor. OpenAI responded seven days later with GPT-5.5 Thinking, priced at $5 per million input and $30 per million output tokens, and immediately claimed the SWE-Bench Verified crown at 88.7%, overtaking Opus 4.7's 87.6%. The gap between the two frontier models is now measured in decimal points, not capability categories.

The competitive timing was not coincidental. Both labs have spent 2026 compressing release cycles, and April's back-to-back launches reflect a market dynamic where sitting at second place for even a month carries real commercial cost. Enterprise procurement teams are increasingly locking in API contracts around benchmark leadership, which means the labs are playing a game where the leaderboard is also the sales deck. That pressure is now structural, not episodic.

The most underreported number from April is not SWE-Bench but MRCR v2. GPT-5.5 Thinking's long-context recall at the 1 million-token mark nearly doubled, rising from 36.6% to 74.0%. That is not an incremental improvement; it is a step-change in what the model can actually do with large codebases, legal corpora, and enterprise document stores. On agentic coding, GPT-5.5 also leads Terminal-Bench 2.0 by a striking 13.3 percentage points, scoring 82.7% against Opus 4.7's 69.4%. Claude holds ground on tool use, with MCP-Atlas at 77.3% versus 75.3%, and edges ahead on GPQA Diamond, 94.2% to 93.6%.

For founders and investors, the pricing picture requires careful reading beyond the rate card. Opus 4.7's new tokenizer can produce up to 35% more tokens for the same input text, meaning real bills can climb even when the listed price holds flat at $5 per million input tokens. GPT-5.5's output pricing at $30 per million tokens is also steep for high-volume inference workloads. Neither model is cheap at scale, and the benchmark war is in danger of obscuring a harder question: which model produces the best output per dollar for a specific production use case, not a leaderboard task.

The next twelve months will determine whether this benchmark convergence translates into genuine commoditisation or whether the labs find new dimensions to differentiate on. Multimodal depth, latency at context scale, and fine-tuning access are the likely next battlegrounds. Founders building on top of either frontier model should be pricing in vendor switching costs now, because the model that leads SWE-Bench today may not lead the metric that matters to their product in six months. The race is accelerating faster than most infrastructure roadmaps can absorb.