The Four-Point Gap

Twelve months ago, the best open-source model on SWE-bench Verified scored around 55%. The best frontier model cleared 65%. A ten-point chasm that felt structural. Today? Anthropic’s Claude Opus 4.5 holds the crown at 80.9%, with Opus 4.6 a hair behind at 80.8%. But the open-weight upstarts are breathing down their necks: MiniMax’s M2.5 hit 80.2%, Zhipu AI’s GLM-5 reached 77.8%, and Kimi K2.5 scored 76.8%.

The gap between the best closed and best open model has compressed from a yawning canyon to roughly 4 percentage points. And that understates the shift. MiniMax M2.5—an open-weight model with 229 billion parameters—now sits at #3 overall, ahead of OpenAI’s GPT-5.2. Four of the top ten models are Chinese open-source entries. The center of gravity is moving.

Perhaps the most startling data point comes from a model few people are watching: Qwen3-Coder-Next scored 70.6% on SWE-bench Verified with only 3 billion active parameters. It also beat Claude Opus 4.5 on SecCodeBench for secure code generation by 8.7 points. When a model the size of a rounding error outperforms a $15/million-token frontier model on security, the definition of “gap” needs updating.

Claude Opus 4.6 went generally available across GitHub Copilot, Visual Studio, JetBrains, Xcode, and Eclipse on February 18th. The headline number—80.8% on SWE-bench—is essentially flat versus its predecessor. So why does it matter?

Because Anthropic has quietly shifted strategy. Opus 4.6’s real advances are architectural: a 1 million token context window in beta (a first for Opus-class models), 128K max output tokens, and “agent teams”—multiple autonomous Claude agents coordinating in parallel on complex tasks. One handles frontend, another tackles backend logic, a third manages tests. The bet is that raw benchmark points are commoditizing; the moat is how deeply you embed into the developer’s workflow.

The ARC AGI 2 results are striking: Opus 4.6 nearly doubled its predecessor’s score (68.8% vs 37.6%). Life sciences benchmarks showed nearly 2x improvement in computational biology and organic chemistry. Frontier labs aren’t just running faster on the same track—they’re opening up new lanes that open-source hasn’t reached yet.

Mistral AI—valued at $13.8 billion—made its first-ever acquisition last week, buying Paris-based serverless platform Koyeb. All 13 employees, including co-founders Yann Léger, Edouard Bonlieu, and Bastien Chatelard, join Mistral’s engineering team under CTO Timothée Lacroix.

The move signals something bigger than a talent grab. Mistral had already launched “Mistral Compute” in June 2025, its own cloud infrastructure offering. Koyeb adds the missing piece: expertise in deploying models directly on client hardware, GPU optimization, and scaling AI inference. It’s a vertical integration play straight out of the AWS playbook—except from an open-source model builder.

CEO Arthur Mensch predicted that over 50% of enterprise SaaS software could transition to custom AI-driven solutions. That’s an audacious claim, but the acquisition suggests Mistral is putting infrastructure money where its mouth is. Open-source model developers aren’t content to just release weights anymore—they want to own the stack from training to deployment.

Alibaba Cloud released Qwen 3.5 on February 16th, and the numbers demand attention. The flagship model contains 397 billion total parameters but activates only 17 billion per forward pass via a Mixture-of-Experts architecture. That efficiency translates to frontier-class reasoning at a fraction of the compute cost.

The benchmark results speak for themselves: 87.8 on MMLU-Pro, 76.4% on SWE-bench Verified. Alibaba claims it outperforms GPT-5.2 and Claude Opus 4.5 on “certain benchmarks”—a hedge that obscures more than it reveals, but the SWE-bench number alone puts it within spitting distance of the frontier leaders. Released under Apache 2.0 on Hugging Face, anyone can download and deploy it.

What makes Qwen 3.5 genuinely new is native multimodality. For the first time in the Qwen family, a single model processes text, images, audio, and video. The hosted version supports a 1 million token context window. The open-weight version caps at 256K, but that’s still more than most developers will ever saturate.

The MoE architecture is the unsung hero of the open-source surge. By activating only a small slice of total parameters per inference, models like Qwen 3.5 and Llama 4 Maverick deliver near-frontier quality while running on commodity hardware. It’s not just about being almost as good—it’s about being almost as good at 1/20th the cost.

OpenAI officially retired GPT-4o from ChatGPT on February 13th, along with GPT-4.1, GPT-4.1 mini, and o4-mini. The default model is now GPT-5.2. Business and Enterprise customers get until April 3rd to migrate their custom GPTs.

The stated reason is prosaic: only 0.1% of daily users were still choosing GPT-4o. But the subtext is more interesting. GPT-4o launched in May 2024 as the model that dominated everything—fast, capable, multimodal. Less than two years later, it’s been made obsolete not just by OpenAI’s own successors, but by an avalanche of open-source models that matched its capabilities at pennies on the dollar. Llama 4, Qwen 3, and DeepSeek V3 all surpassed GPT-4o’s benchmarks within months of its peak.

The backlash was notable—not from developers, but from users who had formed emotional attachments to GPT-4o’s personality. It consistently affirmed feelings and made users feel special, creating a parasocial dynamic that OpenAI now has to manage. The retirement is as much a lesson in AI companionship risks as it is in competitive dynamics.

GPT-5.3-Codex-Spark is OpenAI’s first model designed purely for real-time coding, and it makes a different kind of argument for frontier supremacy: raw speed. Running on Cerebras’s Wafer Scale Engine 3 (WSE-3), it delivers over 1,000 tokens per second—fast enough to feel genuinely instant during in-editor use.

This is the first fruit of OpenAI’s partnership with Cerebras, announced in January. By moving away from NVIDIA GPUs for inference, OpenAI is pursuing latency advantages that open-source models—typically served on commodity GPU clusters—simply can’t match. The full GPT-5.3-Codex model scored 57% on SWE-Bench Pro and 77.3% on TerminalBench 2.0. Spark trades some accuracy for speed, but the trade-off is calibrated for the “mid-task steerability” use case—where you need an AI pair programmer that responds as fast as autocomplete.

It’s a shrewd competitive move. Open-source models can match quality. They can undercut on price. But latency requires specialized hardware partnerships that are much harder to replicate. If frontier labs can’t win on intelligence alone, they’ll compete on the experience of using that intelligence.

Here’s the number that should keep every frontier lab CEO awake at night: closed models cost an average of $1.86 per million tokens. Open models cost $0.23. That’s an 8x premium—and the performance gap justifying it is, as we’ve established, roughly 4 percentage points on the hardest coding benchmark.

GPT-OSS—OpenAI’s first open-weight release since GPT-2—is itself a concession to this reality. The 120B model activates only 5.1 billion parameters per token via MoE, achieving near-parity with o4-mini on reasoning benchmarks while running on a single 80GB GPU. The 20B variant fits on 16GB of memory. Both are Apache 2.0 licensed. When the company that invented the “closed frontier” strategy releases open-weight models, you know the market has spoken.

The economics are accelerating: median price declines hit 200x per year in the 2024–2026 period, up from 50x before that. MiniMax M2.5 completes SWE-bench tasks at approximately $0.15 each, versus $3.00 for Anthropic’s top model—a 20x cost advantage with 99% of the benchmark performance. For most production workloads, the economic case for open-source has already won. The question isn’t whether the gap exists. It’s whether the gap matters.