Table of contents
Open Table of contents
What Actually Shipped
The releases are different in design but converge on the same disruption.
DeepSeek V4 ships in two variants under MIT-licensed open weights on Hugging Face. V4-Pro lists at $1.74 per million input tokens and $3.48 output. V4-Flash drops to roughly $0.14 input / $0.28 output. Both handle a 1-million-token context window. According to MIT Technology Review’s coverage, V4-Pro lines up with the prior closed-frontier generation (Opus 4.6, GPT-5.4, Gemini 3.1) across major benchmarks. Even against Anthropic’s freshly released Opus 4.7-which lifted SWE-bench Verified from 80.8% to 87.6% and SWE-bench Pro from 53.4% to 64.3%-V4-Pro remains in the same performance band on most workloads at roughly one-tenth the per-token output cost. Over 90% of 85 surveyed developers ranked V4-Pro among their top choices for coding.
The architectural story matters as much as the pricing. V4-Pro is a 1.6-trillion-parameter MoE with 49B active per token; V4-Flash is 284B total with 13B active. The 1M-context capability without quality collapse comes from a hybrid-attention design published in the V4 technical report that combines two compression schemes: Compressed Sparse Attention (dynamic KV-cache compression plus DeepSeek Sparse Attention to sparsify attention matrices) and Heavily Compressed Attention (consolidating KV entries across token groups into single compressed representations). The result is 73% fewer per-token inference FLOPs and a 90% reduction in KV cache memory versus V3.2-with quality preserved because the heaviest compression is applied where attention contributes least to output. At full 1M context, V4-Pro runs on 27% of V3.2’s compute and 10% of its memory; V4-Flash hits 10% compute and 7% memory. That’s why long-context inference economics no longer collapse at scale-and because the design is published and openly licensed, the efficiency wins propagate through the rest of the open-weight ecosystem rather than staying behind a vendor’s API.
Kimi K2.6, from Moonshot AI, takes the same direction with different choices. It’s a 1-trillion-parameter mixture-of-experts model with 32B active parameters and 384 experts per layer, released under a Modified MIT license. The official Moonshot API prices it at $0.60 input / $2.50 output per million tokens. Context window is 262K with 16K max output-shorter than DeepSeek V4, but Kimi’s edge is agent orchestration: documented agent swarms scaling to 300 concurrent sub-agents across 4,000 coordinated steps, with SWE-Bench Pro at 58.6% and HLE-Full with tools at 54.0%. That positioned Kimi ahead of the prior closed frontier (Opus 4.6 at 53.4% on SWE-Bench Pro) at release; Opus 4.7’s 64.3% has since stepped ahead on this specific benchmark, but Kimi’s pricing remains roughly 10x lower on output and the agent-orchestration capability is now broadly competitive with the closed leaders.
Two different architectures, same conclusion: open weights are no longer the cheap-and-good-enough alternative. They’re competitive on the tasks enterprises are actually buying AI for.
The Cost and Capability Picture
The pricing gap is large enough to change deployment decisions, not just budget lines.
Compare list pricing per million tokens (input / output) against Anthropic’s current lineup:
- Kimi K2.6 (Moonshot): $0.60 / $2.50
- DeepSeek V4-Pro: $1.74 / $3.48
- Claude Haiku 4.5: $1 / $5
- Claude Sonnet 4.6: $3 / $15
- Claude Opus 4.7: $5 / $25 (rate unchanged from 4.6, but the new tokenizer can produce up to 35% more tokens for the same input-so effective bills can rise even though the rate card didn’t)
Sonnet 4.6 is already a meaningful step down from Opus pricing while matching it on most enterprise workloads. DeepSeek V4-Pro undercuts Sonnet by roughly 5x on output. Kimi K2.6 undercuts Opus 4.7 by a factor of 10 on output-and that’s before the 4.7 tokenizer change is factored in. At enterprise volumes, that’s not a margin improvement-it’s a different cost structure.
This matters specifically because 80-90% of AI spend lives in production inference, not training or experimentation. A 5–10x reduction in per-token cost compounds aggressively when an agentic workflow burns millions of tokens per task. The same workload that constrains a budget on Opus 4.7 becomes routine on V4-Pro.
The capability story has caught up with the price story. That’s the new fact.
Why the Lock-In Math Changed
Open-weight licensing has been available throughout 2025, but the trade-off was real: you accepted a performance gap to escape vendor dependency. That trade is what’s gone.
There are now three deployment options that produce viable enterprise AI:
- Hosted closed APIs (Anthropic, OpenAI, Google) - strongest tooling, predictable contracts, mature support, premium pricing
- Hosted open-weight APIs (DeepSeek, Moonshot, third-party providers) - frontier-class capability, much lower per-token cost, dependency on a single provider’s terms
- Self-hosted open weights - full control over inference economics and data flow, requires infrastructure capability and operational maturity
The middle option used to be a compromise. It isn’t anymore. And the third option, which most enterprises dismissed as too operationally heavy, becomes credible when the model in question matches the closed frontier.
This reshapes vendor risk in three concrete ways:
Pricing leverage. When equivalent capability exists at one-fifth the price, your closed-API contract is no longer a take-it-or-leave-it. Renewal conversations look different when both sides know the alternative is real.
Continuity insurance. A model whose weights you can download and run can’t be deprecated, repriced, rate-limited, or reshaped by a vendor’s changing business model. For workloads that matter, that’s a different category of resilience.
Architecture flexibility. With open weights, you decide where inference runs. That choice has implications for data residency, latency, regulatory posture, and unit economics that closed APIs don’t let you optimize.
None of this means closed-frontier vendors are losing. Anthropic’s agentic stack and OpenAI’s tooling depth still justify premium pricing for the workloads that need them. But “premium for everything by default” is no longer a defensible architecture.
Open Weights as the Sustainability Bet
The vendor-math case for open weights is immediate and quantifiable. The sustainability case is structural-and it’s why the trajectory matters beyond this quarter’s budget.
By 2026, roughly 63% of frontier-model lifecycle energy is inference, 37% training-a complete inversion from two years ago. The per-query efficiency of the model you actually deploy now matters far more than the training-stage choices that get the headlines. Three open-weight properties compound there in ways closed APIs cannot match:
Architectural efficiency propagates. DeepSeek’s hybrid-attention paper is public. Any team can adapt it. V4’s 73% FLOP reduction and 90% KV-cache reduction at 1M context aren’t proprietary secrets-they’re a published recipe that the rest of the open-weight ecosystem will absorb in the next release cycle. With closed models, you only get the efficiency the vendor chooses to pass on, and you can’t verify what’s actually under the hood.
Deployment location is yours to choose. Inference emissions per query vary 5–10x between coal-heavy and renewable-heavy grids. With closed APIs you take whatever grid mix the provider runs. With open weights you can route inference through low-carbon-grid providers (Pacific Northwest, Iceland, Quebec, Scandinavia)-the same lever explored in our GreenOps playbook for cloud workloads more broadly. The carbon savings are real and they accrue to the workloads you actually run, not to a vendor’s marketing report.
Right-sized models per workload. A 70B-class FP8 model uses dramatically less energy per query than a 1.6T MoE for many use cases. Open weights make it practical to distill, fine-tune, or quantize down to the smallest adequate model for a given task-not just default to frontier capability for everything. Closed APIs price by tier; open weights let you size by workload.
None of this is automatic. Self-hosted inference can be less efficient than hyperscaler operations if you don’t know what you’re doing, and the operational maturity required is real. But the structural option exists in open weights and doesn’t in closed APIs-and it compounds. As AI absolute energy use continues climbing 38% year-over-year while per-query efficiency improves, the choice between “vendor optimizes when it suits them” and “we optimize directly” stops being a small decision.
The Geopolitical Asterisk
Honest analysis has to account for the regulatory exposure.
Both releases come from Chinese labs. DeepSeek V4 is the first model explicitly optimized for Huawei’s Ascend chips for inference, with partial training also adapted to domestic Chinese hardware. US lawmakers escalated calls in April 2026 to add DeepSeek and several Chinese AI labs to the Commerce Department’s Entity List, with the Huawei training claim and V4’s open release cited as triggers.
The practical implications for enterprise buyers:
- Already-downloaded weights are not retroactively illegal, but redistribution and commercial deployment in restricted sectors enters a gray zone legal teams typically won’t approve for production.
- Hosted APIs from Chinese providers carry data-flow exposure that healthcare, finance, defense, and federal-adjacent buyers cannot absorb regardless of price.
- Third-party hosting of open weights (US or EU providers running DeepSeek V4 or Kimi K2.6 on their own infrastructure) is the cleanest path for regulated buyers, but the hosting market is still maturing and the contractual specifics matter.
This isn’t a reason to ignore the releases. It’s a reason to make explicit decisions about which workloads can use which deployment path. For unregulated internal tooling, the cost savings are immediate. For regulated production systems, the same model may not be deployable for months and may never be deployable from the original provider.
What Enterprises Should Do
The April 2026 releases don’t demand a rip-and-replace. They demand re-evaluation.
Re-baseline your inference economics. Take the three highest-volume AI workloads and model them against open-weight pricing, including realistic infrastructure costs for self-hosted scenarios. If the gap is 5x or larger, that’s a budget conversation that should happen this quarter.
Decouple model choice from architecture. If switching providers requires rewriting application logic, that’s lock-in dressed as integration. Move toward provider-agnostic orchestration-standard chat completion interfaces, prompt management outside the application, evaluation harnesses that can swap models without code changes.
Run a real evaluation, not a benchmark scan. Published benchmarks tell you a model can perform on standardized tasks. They don’t tell you whether it performs on yours. Take a representative workload sample (50–200 examples) and run it across DeepSeek V4-Pro, Kimi K2.6, and your current closed model. Compare quality, latency, and cost honestly.
Map your regulatory boundaries before you deploy. Decide which workloads can use Chinese-hosted APIs, which require third-party hosting, and which stay on closed Western providers regardless of price. Get those decisions in writing before someone makes them ad hoc.
Factor grid-mix into deployment decisions. Inference now dominates AI lifecycle energy, and emissions per query can vary 5–10x by region. Where workloads are flexible, route open-weight inference through low-carbon-grid providers. The cost savings often correlate with the carbon savings.
Renegotiate from a stronger position. If a major closed-API contract is up for renewal in the next two quarters, the pricing landscape has changed underneath that contract. Use it.
The Bigger Signal
Open-weight frontier parity isn’t a one-time event. DeepSeek V3 hinted at this trajectory in late 2024. R1 confirmed it in early 2025. V4 and Kimi K2.6, four days apart, make it routine. The compounding rate of capability-per-dollar in open weights has now exceeded the rate at which the closed frontier is moving forward.
For enterprises, the strategic implication is straightforward: AI cost structures that assumed permanent dependency on one or two closed-API vendors need to be revisited. The companies that benefit most won’t be the ones that switch wholesale to open weights. They’ll be the ones who build architectures that let them route each workload to the right model-closed when capability or compliance demands it, open when economics and sustainability allow it-and who do that routing deliberately rather than by inertia.
The deeper shift is that open weights are increasingly the more sustainable substrate for production AI. Architectural efficiency propagates rather than staying locked behind APIs. Deployment location becomes a lever for carbon, not just latency. Right-sizing replaces tier-buying. None of that is guaranteed by the licensing alone-it’s a structural option that has to be exercised. But it’s an option that closed-API consumption fundamentally doesn’t provide.
The vendor math changed in April. The architecture decisions to match it-on cost, on resilience, and on emissions-are still ahead.
Sources
- MIT Technology Review, “Why DeepSeek’s V4 matters” (April 24, 2026)
- Winbuzzer, “DeepSeek V4 Ships 1M Context, Open-Weights” (April 27, 2026)
- OpenRouter, “Kimi K2.6 - API Pricing & Benchmarks”
- Codersera, “Kimi K2.6 Complete Guide (2026): Benchmarks, Pricing, Agent Swarms”
- Anthropic, “Introducing Claude Opus 4.7” (April 16, 2026)
- Finout, “Claude Opus 4.7 Pricing 2026: The Real Cost Story Behind the ‘Unchanged’ Price Tag”
- Hugging Face, “DeepSeek V4 Technical Report Summary” - hybrid attention architecture, 73% FLOP / 90% KV-cache reduction
- NVIDIA Technical Blog, “Build with DeepSeek V4 Using NVIDIA Blackwell” - Compressed Sparse Attention and Heavily Compressed Attention details
- Digital Applied, “AI Model Sustainability Report 2026: Energy Use Data” - inference vs training lifecycle energy, grid-mix emissions delta