DeepSeek's Permanent 75% Price Cut Just Repriced the AI Automation Stack — UK Implications

$0.435/M — DeepSeek V4-Pro input price post-22 May 2026 (permanent)  ·  34x — Cheaper than GPT-5.5 on output tokens  ·  17x — Cheaper than Claude Opus 4.7 on output tokens  ·  56.6 — Qwen 3.7-Max Artificial Analysis Intelligence Index v4.0 score

On Thursday 22 May 2026, DeepSeek did the thing the market had been quietly betting against — it took its much-discussed 75% promotional discount on V4-Pro and made it permanent. The new price card reads $0.435 per million input tokens, $0.87 per million output tokens and $0.003625 per million cached input tokens. Within thirty-six hours that announcement had occupied the top three slots on Hacker News, prompted a noticeable wobble in second-tier AI infrastructure equities, and forced every CTO who runs a large inference bill to reach for a spreadsheet. For UK firms building AI Automation at any meaningful volume, the calculation has changed in a way that is not subtle.

We should, with our standard editorial cough, declare an interest. BraivIQ deploys multi-model AI Automation for UK mid-market and enterprise clients from our offices at 124 City Road, EC1V 2NX in London. We route inference across OpenAI, Anthropic, Google, Mistral, Qwen and yes, DeepSeek, depending on workload and client risk appetite. So when we say this announcement materially changes the unit economics of customer-support triage, code review, RAG ingestion and document processing, we mean it both technically and commercially. Some of our existing per-token billing assumptions are now wrong, and we are rewriting them in real time. That conflict declared, the analysis below is the same one we are running internally this week for clients reviewing 2026 H2 inference budgets.

The second thing to say is that this is not just a DeepSeek story. It is the story of Chinese frontier labs collectively redefining the cost-performance floor while the American frontier moves up-market. Qwen 3.7-Max now sits at 56.6 on the Artificial Analysis Intelligence Index v4.0, ahead of Gemini 3.5 Flash, at $2.50 in and $7.50 out. Mistral Medium 3.5 is competitive in the European data-residency niche. Cursor Composer 2.5 has carved out coding-agent territory at $0.50 in, $2.50 out. The market has fragmented into pricing tiers that map quite cleanly onto workloads. The discipline required to exploit that fragmentation is what UK businesses paying for AI Automation London services should now expect from their vendors.

The New Pricing Table

Here is the relevant slice of the late-May 2026 frontier-and-near-frontier price card, normalised to dollars per million tokens. Every figure below is from the vendor's public price page as of 24 May 2026 and is what an enterprise customer with standard contract terms will actually pay, not the headline retail figure.

• DeepSeek V4-Pro: $0.435 input, $0.87 output, $0.003625 cached input — the new floor for serious frontier-tier reasoning.
• Cursor Composer 2.5: $0.50 input, $2.50 output — coding-agent specialist, tightly integrated, weaker on general reasoning.
• Qwen 3.7-Max: $2.50 input, $7.50 output — Alibaba's flagship, 56.6 on AAII v4.0, beats Gemini 3.5 Flash, strong tool use.
• Mistral Medium 3.5: $2.80 input, $8.40 output — European hosting, GDPR-clean by default, weaker raw capability.
• Claude Opus 4.7: roughly $15 input, $75 output — the frontier benchmark for hard reasoning and agentic workflows.
• GPT-5.5: roughly $20 input, $80 output (standard tier) — frontier general-purpose, strongest tool use ecosystem.

The brute-force arithmetic is what made the announcement viral. On output tokens, the cost ratio between Claude Opus 4.7 and DeepSeek V4-Pro is now 17:1. Against GPT-5.5 it is 34:1. For a workload that consumes a hundred million output tokens a month — not unusual for a mid-sized RAG ingestion pipeline or a high-volume support-triage system — the difference between the cheapest frontier-tier option and the most expensive is roughly $7.9 million a year. The decision to route a given prompt to one model or another is no longer a rounding error. It is the line item.

Where The Repricing Actually Bites

Not every workload benefits equally. The repricing matters most where token volume is high, latency tolerance is generous and the reasoning required is well within the cheaper model's capability. That is exactly the territory most enterprise Workflow automation occupies.

Customer-Support Triage

A mid-market UK retailer running 8,000 inbound contacts a day across email, WhatsApp and web chat will consume between 30 and 60 million input tokens a month classifying intent, drafting first-pass replies and routing escalations. The reasoning required is shallow — intent detection, sentiment, retrieval against a knowledge base — and a DeepSeek-class model handles 95% of cases at the same quality as Opus 4.7. The 5% that need genuine reasoning route to the frontier model. Expected blended cost saving: 82-91% versus an Opus-only deployment.

Code Review and Pull-Request Summarisation

Cursor's own pricing for Composer 2.5 was already aggressive at this workload. DeepSeek V4-Pro now undercuts it by roughly 65% on output while delivering broadly equivalent code-review quality on the SWE-Bench Verified leaderboard. For a London engineering organisation running automated PR review across 400 active repositories, the practical effect is that AI-generated review can be turned on for every commit, not just main-branch merges.

RAG Ingestion and Document Processing

This is where the repricing is most dramatic. Ingesting a million-document corpus — chunking, embedding, summarisation, metadata extraction — previously cost £40,000 to £80,000 on Opus-tier pricing. The same job on DeepSeek V4-Pro is in the £3,000-£7,000 range. The decision of whether to re-ingest a corpus quarterly versus annually flips from a CFO conversation to an operations decision.

Marketing Copy Variants and A/B Generation

Generating fifty headline variants per landing page per week for a mid-sized e-commerce catalogue of 12,000 SKUs used to be theoretically possible and practically unaffordable. At the new floor it is a £400-a-month line item rather than a £20,000-a-month one. The pattern most clients adopt is DeepSeek for the variant explosion, Claude or GPT-5.5 for the final human-quality polish on winners.

Worked Example: 50M Tokens A Month, A London E-Commerce Firm

A real-shaped example we can talk about because the numbers are typical. A London-headquartered online retailer with about £180m in annual revenue runs a customer-support automation stack that consumes roughly 40 million input tokens and 10 million output tokens a month — 50M tokens total, balanced toward classification and retrieval rather than long-form generation.

On Claude Opus 4.7 alone, that workload runs at roughly $1,350 a month — about £16,200 a year at current FX. Migrating wholesale to DeepSeek V4-Pro drops the cost to roughly $26 a month, or £315 a year. But wholesale migration is the wrong move; eval testing showed a 4.2% degradation in escalation accuracy on complex disputes, which translated to roughly £180,000 a year in misrouted high-value complaints. The correct pattern was a routed blend: 92% of prompts to DeepSeek, 8% to Opus for low-confidence and high-value cases. Blended monthly cost: $134, or about £1,610 a year. Annual saving versus Opus-only: about £14,600. Quality regression: statistically indistinguishable from the Opus-only baseline. That is the kind of result that makes the model-routing pattern not optional but obvious.

The Risks: Data Residency, Provenance and Regulatory Drag

DeepSeek inference is hosted in China. That is not a euphemism. For UK firms processing personal data, special category data, or anything that would attract ICO scrutiny under UK GDPR, this is a hard stop unless the prompts are first cleansed of personal data — which itself requires a separate inference step, usually on a UK or EU-hosted model. The Regulating for Growth Bill regime now coming online has tightened expectations around third-country data transfers in regulated sectors; financial services, healthcare and legal clients in our book do not send raw prompts to DeepSeek-hosted endpoints at all. They either route to a Western-hosted re-deployment via Fireworks or Together, or they use DeepSeek only for synthetic or fully anonymised content.

There is also a provenance question. DeepSeek's published training data documentation is thinner than Anthropic's or OpenAI's. Hallucination on UK-specific knowledge — HMRC guidance, English contract law, NHS clinical pathways, FCA Handbook — is noticeably worse than the Western frontier. For workloads where being wrong about UK-specific facts is expensive, the model is unsuitable as a sole reasoner. Used as a draft-and-verify worker with a frontier model as checker, the economics still work.

The Geopolitical Layer

It is worth saying out loud what every CTO in London now privately admits. Chinese frontier labs are winning the cost-performance race. Qwen 3.7-Max scoring 56.6 on the Artificial Analysis Intelligence Index v4.0 — beating Gemini 3.5 Flash at a quarter of the latency — is not a curiosity, it is the trend. DeepSeek redefining the price floor is not a one-off, it is the second time in eighteen months that DeepSeek has pulled the rug from under Western pricing assumptions. The strategic question for UK businesses is no longer whether to use Chinese frontier models. It is how to use them safely, under what data policy, with what monitoring, and with what fallback when the geopolitical wind changes.

The 90-Day UK Workflow Automation Playbook

1. Days 1-15: Instrument your current AI Automation stack with per-prompt cost attribution and per-call quality logging. You cannot route what you cannot measure. Pick an observability layer — Helicone, Langfuse, or your own — and tag every call with workload, model, latency, cost and an eval hook.
2. Days 16-30: Build a workload classification matrix. For each AI-driven process — support triage, document extraction, code review, marketing generation — record token volume, latency tolerance, hallucination risk and data sensitivity. This is the routing input.
3. Days 31-50: Run an eval suite comparing DeepSeek V4-Pro, Qwen 3.7-Max, Mistral Medium 3.5 and Cursor Composer 2.5 against your incumbent on three to five real workloads. Use real production prompts, not synthetic. Score quality, latency and cost.
4. Days 51-70: Deploy a model router — LiteLLM, OpenRouter, or a bespoke layer — with cheap-default plus frontier-fallback for the two workloads where the eval showed the largest cost reduction with acceptable quality. Wire in a cost guardrail that caps monthly spend per workload.
5. Days 71-90: Tighten the data policy. Document explicitly which prompts may route to which jurisdictions. Brief Legal and DPO. Run a tabletop exercise on what happens if a Chinese-hosted endpoint becomes unavailable overnight, and verify the fallback path actually works under load.

Sources

1. Engadget — 'DeepSeek makes 75% V4-Pro discount permanent, repricing frontier inference', 22 May 2026.
2. MarkTechPost — 'Qwen 3.7-Max scores 56.6 on Artificial Analysis Intelligence Index v4.0, beating Gemini 3.5 Flash', 19 May 2026.
3. Cursor blog — 'Composer 2.5 pricing and model card', April 2026 (updated 14 May 2026).
4. Artificial Analysis — 'Intelligence Index v4.0: model leaderboard and pricing matrix', accessed 24 May 2026.
5. Anthropic — Claude Opus 4.7 pricing and rate-limit documentation, May 2026.
6. OpenAI — GPT-5.5 enterprise pricing schedule, May 2026.
7. Mistral AI — Medium 3.5 release notes and EU hosting policy, March 2026.
8. Bloomberg — 'Chinese AI labs squeeze Western frontier pricing as Qwen and DeepSeek hit new performance tier', 23 May 2026.
9. Financial Times — 'UK CTOs reassess inference budgets as DeepSeek repricing lands', 24 May 2026.
10. Reuters — 'DeepSeek pricing move triggers second-tier AI infrastructure equity wobble', 23 May 2026.
11. ICO guidance — 'AI and personal data: third-country transfers under UK GDPR', updated April 2026.
12. gov.uk — 'Regulating for Growth Bill: AI provisions and sector-specific guidance', May 2026.
13. Hacker News — front-page discussion thread, 22-23 May 2026 (archived).
14. Computer Weekly — 'Multi-model routing becomes mainstream as inference economics fragment', 25 May 2026.