The open-weights rebound: capability parity at one-tenth the price
DeepSeek V4 under MIT, GLM-5.1 at $0.18/M, Kimi K2.6 at 256K context, Llama 4 Maverick. The open-weight frontier is now within a few SWE-bench points of closed flagships at one-tenth the input cost. The structural implications run deeper than pricing.
What changed in six months
DeepSeek shipped V4 on April 24 — Pro at 1.6T total / 49B active, Flash at 284B total / 13B active, both at 1M context, both MIT-licensed. The flagship is competitive with closed models on agentic and reasoning benchmarks. Flash is at $0.14 per million input tokens. That is the floor for the open-weight frontier, and every other lab is now pricing relative to that number.
Z.ai's GLM-5.1 sits at $0.18/M. Moonshot's Kimi K2.6 is a 1T-parameter MoE with 32B active and a 256K context. Llama 4 Maverick — the open-weights lineage Meta has continued investing in despite the strategic tax of doing so — is in the same capability tier.
The capability gap closed without the lead closing
It's important to separate two claims:
- The capability gap closed. Open-weight models are within a few percentage points of closed flagships on SWE-bench Verified, MMLU, and most general benchmarks. For most deployed tasks, the gap is no longer the bottleneck.
- The frontier-lead is still real. Mythos Preview's 93.9% SWE-bench and TLO partial-clear are not yet matched by any open-weight model. The frontier moves; the open-weight tier catches up about two to three months later.
Both can be true. The market's reaction is what matters. For 80% of enterprise tasks, "good enough at one-tenth the cost" wins over "best in class at full price." For the remaining 20% — frontier-research, regulated capability gates, agentic long-horizon — the closed labs keep their premium.
Why the gap closed
Three structural reasons:
- The recipe leaked. Mixture-of-experts at trillion-parameter scale with sparse activation is no longer secret. Pre-training data curation is harder to copy, but the architecture is now common knowledge.
- The compute commoditized. NVIDIA's grip loosened, hyperscaler procurement diversified, Cerebras IPO'd on the back of OpenAI. Compute is still expensive, but it's no longer a moat.
- National-policy motives. DeepSeek, Z.ai, and Moonshot are operating in a strategic context where open-weight releases serve national-AI-ecosystem goals. The economics are subsidized in ways that look uncompetitive from a pure-private-firm perspective.
What enterprises should do now
The decision framework changed. Six months ago the question was "closed flagship or fine-tuned open model?" — and the answer for serious workloads was almost always the closed flagship. Now the question splits into three:
For routine inference: open-weight, self-hosted or rented, at one-tenth the cost. For frontier capability needs: closed-flagship, paid. For regulated or capability-gated tasks: whichever lab has the safety attestation your compliance team requires.
Few enterprises actually need the frontier tier. Most need the routine tier and overpay. The 10× price differential makes "audit your inference workload by capability requirement" the highest-ROI exercise of Q3 2026.
The closed-lab response
The closed labs have three available responses, none of them clean:
- Price-match. Cuts margin, doesn't win on capability differentiation, signals weakness. OpenAI's existing tiering (mini, standard, pro) is the start of this.
- Lean harder into frontier-only differentiation. Anthropic's Mythos hold is a version of this — make the disclosure that you have capability the open tier doesn't, and accept that you're priced as a premium product.
- Open-weight a tier. The structural option none of the closed labs have taken. OpenAI's gpt-oss release was a half-step in this direction. Whether any closed lab open-weights a flagship in 2026 is the question of the year.
The 2027 question
Two trajectories are possible:
- The gap stays constant. Open-weights catch up to last quarter's frontier on a rolling basis. Closed labs keep a quarter of differentiation. Both ecosystems grow.
- The gap closes entirely. Open-weights match the frontier in real time. Closed labs become services and tooling companies on top of commodity capability. The price floor collapses further.
The second trajectory is what closed labs are betting against. The first is what's actually happening. Whether trajectory two arrives depends almost entirely on what DeepSeek V5, GLM-6, and Llama 5 ship — and whether the next frontier-only-disclosure (Mythos's successor) introduces a capability that genuinely doesn't replicate at open-weight scale.
The honest read
The open-weights rebound is not a hypothesis anymore. It's the current market state. The question is whether it persists as a rolling-catch-up dynamic or accelerates into full parity. Either way, the era of closed-flagship-as-default for enterprise inference ended in May 2026, even if most procurement teams haven't noticed yet.