4 Models. 12 Days. The Open-Source AI Price War Just Ended — And China Won

4 Models. 12 Days. The Open-Source AI Price War Just Ended — And China Won

Last month, four Chinese AI labs dropped competing open-weight coding models in a 12-day window. Each one benchmarks within striking distance of Claude Opus 4.6. Each one costs a fraction of the price. The open-source AI "catch-up" narrative? It's dead.

Here's what happened — and what it means if you're building anything with AI right now.

The April 2026 Tsunami

Between April 7 and April 24, 2026, the following models went live:

Date	Model	Lab	Key Stat
Apr 7	GLM-5.1	Z.AI (Zhipu)	First open-weight #1 on SWE-Bench Pro
Apr 20	Kimi K2.6	Moonshot AI	Beat GPT-5.4 on SWE-Bench Pro
Apr 24	DeepSeek V4	DeepSeek	93.5% LiveCodeBench — highest of any model
Mar 18	MiniMax M2.7	MiniMax	10B active params, 56.22% SWE-Pro

That's not a slow evolution. That's a tectonic shift.

Kimi K2.6: The Open-Weight King for Coding

Moonshot AI's Kimi K2.6 is a 1-trillion-parameter MoE model with 32B active parameters. It scored 58.6% on SWE-Bench Pro — beating GPT-5.4 (57.7%) and Claude Opus 4.6 (53.4%) on the hardest verified coding benchmark.

The number that matters for production teams: $0.60 per million input tokens. That's roughly 8x cheaper than Opus 4.7 at $5/M.

K2.6 also handles autonomous coding sessions over 12 hours and coordinates up to 300 sub-agents across 4,000+ tool calls. This isn't a model that wrote a function — it's a runtime architecture for shipping code.

Where it shines: Long-running agent loops, multi-file refactors, polyglot projects (Rust, Go, Python).

Where it doesn't: The 256K context window is fine for most work, but monorepo analysis across hundreds of files will hit the ceiling.

DeepSeek V4 Pro Max: The Cost Shock

DeepSeek V4 landed on April 24 and immediately redefined what "frontier-class" costs.

The numbers:

• V4-Pro-Max: 1.6T total params, 49B active, 1M context, MIT license
• LiveCodeBench: 93.5% (highest of any model, period)
• SWE-bench Verified: 80.6% — 0.2 points behind Claude Opus 4.6
• Codeforces Elo: 3,206 (ahead of GPT-5.4's 3,168)
• Output cost: $3.48/M vs Opus 4.6's $75/M

That's a 21x cost reduction at near-identical coding benchmark performance.

The Flash variant is even more absurd: $0.14/M input tokens. For high-volume batch coding workflows, that's a different economic category entirely.

NIST's CAISI evaluation confirmed V4 is "the most capable PRC AI model to date" — though they noted it still lags the Western frontier by about 8 months on certain reasoning tasks. Fair enough. At 1/21st the price, I'll take that trade.

GLM-5.1: MIT License, Frontier Performance

Z.AI's GLM-5.1 made history on April 7 as the first open-weight model to top SWE-Bench Pro at 58.4%. It held that #1 spot for nine days until Claude Opus 4.7 reclaimed it.

754 billion parameters. MoE architecture. MIT license — meaning fine-tune, self-host, redistribute, no questions asked.

On real-world agentic tasks, it scores 1,535 on the GDPval-AA leaderboard, placing it third globally for agentic web development. The Code Arena Elo of 1,530 reflects actual developer preference in head-to-head comparisons.

Pricing: Starting from $0.60/M input tokens (API) or free if you self-host.

The MIT license is the real story here. Enterprise teams that need full weights access without commercial restrictions now have a legitimate frontier-class option.

MiniMax M2.7: The Efficiency Play

MiniMax M2.7 is the dark horse. With only 10B active parameters out of 230B total, it scores 56.22% on SWE-Bench Pro — 94% of GLM-5.1's performance at roughly one-fifth the per-token cost.

At $0.30/M input tokens, it's the cheapest option for high-throughput coding workloads. MiniMax also reported that an internal version of M2.7 autonomously optimized a programming scaffold over 100+ rounds, achieving a 30% performance improvement through self-evolution.

The catch: M2.7's license shifted from MIT (M2/M2.5) to non-commercial. For research, prototyping, and internal tooling, it's excellent. For production products, you'll need an enterprise agreement.

What About Qwen?

Alibaba's Qwen family continues to dominate specific niches:

• Qwen 3.6 Plus leads Terminal-Bench 2.0 at 61.6% with a 1M token context window — the only model in this group that can handle monorepo-scale analysis
• Qwen3-Coder-Next (released May 19) is a lightweight 80B MoE with only 3B active, purpose-built for coding agents at $0.11/M input
• Qwen 3.5-397B-A17B scored 83.6 on LiveCodeBench v6 with only 17B active parameters

Qwen's strategy is clear: own the cost-efficient, open-weight middle ground while closing the gap on flagship performance.

Brand New: Ring-2.6-1T

Just this week, InclusionAI (under Ant Group) released Ring-2.6-1T — a 1-trillion-parameter thinking model under MIT license. It's built for agent workflows, coding, and long-horizon reasoning. Early benchmarks show it surpassing GPT-5.4 and Gemini 3.1 Pro on certain tasks. This one's worth watching.

The Real Story: Price-Per-Quality Has Collapsed

Here's the comparison that changes your budget conversation:

Model	Input $/M	Output $/M	SWE-bench Verified	LiveCodeBench
Claude Opus 4.6	$15.00	$75.00	80.8%	88.8
GPT-5.5	$5.00	$30.00	88.7%	—
DeepSeek V4 Pro Max	$0.55	$3.48	80.6%	93.5
Kimi K2.6	$0.60	$2.50	~72% (Tier A)	89.6
GLM-5.1	$0.60	$2.00	~74%	—
MiniMax M2.7	$0.30	$1.20	78%	—

DeepSeek V4 Pro Max scores 80.6% on SWE-bench Verified at $3.48/M output. Claude Opus 4.6 scores 80.8% at $75/M output. That's a 21x price gap for 0.2 benchmark points.

For teams running thousands of agentic coding tasks per day, this isn't incremental savings. It's a structural shift in what's economically feasible.

What This Means for Developers

1. The "open-source is two years behind" argument is empirically wrong. GLM-5.1 topped SWE-Bench Pro. DeepSeek V4 matches Opus 4.6 on SWE-bench Verified. Kimi K2.6 beat GPT-5.4. All three are built without Nvidia hardware.

2. The harness matters more than the model. The gap between raw model capability and production outcomes is now determined by your eval discipline, tooling, and reliability instrumentation. Pick the model that fits your workload, then spend more attention on the layer above it.

3. Self-hosting is a real option. GLM-5.1 (MIT), DeepSeek V4 (MIT), Kimi K2.6 (modified MIT), and Qwen3.6-27B (Apache 2.0) all support self-hosting. For teams with latency requirements or data residency needs, this changes the architecture conversation.

4. The pricing gap is 5-25x and widening. Chinese-stack inference runs 5-25x cheaper at frontier capability tiers. As DeepSeek's V4-Flash at $0.14/M input demonstrates, the floor keeps dropping.

Coming Next

• Kimi K3 — teased March 28, expected Q3 2026. 1M context, 3-4T total parameters
• DeepSeek V5 — aggressive cadence continues
• Qwen 4 — Alibaba's annual cycle suggests Q3/Q4 announcement
• GPT-6 — late May to early July 2026, with long-term memory as the headline feature

The next 90 days will be the most competitive window in AI history. If you're not evaluating these models against your actual workloads, you're making architecture decisions based on last year's pricing.

---

Data sourced from LLM Stats, OpenRouter API, BenchLM.ai, Artificial Analysis, NIST CAISI evaluations, and provider documentation as of May 20, 2026.