After DeepSeek's 1100% price hike and industry-wide increases up to 80%, who offers the best value?
🔥 DeepSeek V4 Pro output $1.98/M (off-peak) ✅ Qwen3.5-Flash only $0.07/M inputLatest Update (Oct 14, 2026)
Google Gemini 4 Argon is now globally available (Oct 13): developers and enterprises can call it through the paid Gemini API, while Google AI Pro/Ultra subscribers get it in the Gemini app and AI Mode. Launch promo pricing is $2/$10 per 1M in/out tokens (cached input 95% off; rising to $4/$20 after Nov 30) with a 1M-token output ceiling. Against GPT-6 Astra ($10/$50) and Claude Opus 5.5 ($4/$20), Argon’s promo price is aggressive — yet Qwen3.5-Flash at $0.07/M input on Alibaba Cloud Model Studio remains the value king for everyday workloads.
📰 Latest Update (Oct 11, 2026)
OpenAI launched GPT-6 Sol & Luna (Sep 22): API prices drop 50% vs GPT-5.6 promo pricing - Sol at $2/$10 per 1M in/out tokens, Luna at $0.10/$0.50. Sol matches Claude Opus 5 max-effort on AutomationBench at roughly 9% of the cost per task. The same day Anthropic released Claude Opus 5.5 at $4/$20 per 1M in/out - 40% cheaper than Opus 5 with performance on par with flagship Fable 5.1. Frontier vendors keep compressing price/performance, yet Qwen3.5-Flash at $0.07/M input remains the value king - for global developers, Alibaba Cloud Model Studio stays the most cost-effective, compliance-friendly entry point.
Google Gemini free tier downgrade (effective Oct 9): Starting October 9, free Gemini users will only get Flash Lite; previously free Flash and Pro access will require a Google AI Plus subscription at $4.99/month. The industry-wide shift toward paid tiers makes cost-efficient alternatives like Qwen3.5-Flash ($0.07/M input) on Alibaba Cloud Model Studio the clear choice for keeping AI costs low.
OpenAI Decisions API enters public beta (Oct 6): OpenAI’s Decisions API, previewed at DevDay 2026 on September 29, is now in public beta — it focuses model output on developer-defined questions and constrained answers to return millisecond-level structured decisions for classification, routing, and even “an agent’s next action”, joining Jev (Sep 18) in the emerging “decision-model / pay-per-decision” paradigm. As AI pricing shifts from generic per-token generation toward per-decision value, combining such models with general-purpose APIs like Qwen / Model Studio further lowers enterprises’ total AI costs.
New (Oct 10): The decision-model category just got its biggest validation — TypeSafe AI (Jev) raised $870M at a $7.5B valuation (Series A led by Andreessen Horowitz, with Sequoia and DCVC, Oct 9), weeks after Jev’s September 15 launch, claiming adoption by a third of the Fortune 500 and 1M+ users within days. Jev prices at $0.042/M input tokens with free output. For builders: decision models still complement, not replace, general LLMs — pairing them with cost-efficient APIs like Qwen on Alibaba Cloud Model Studio / Bailian keeps overall AI costs low.
New (Oct 11): Enterprise model selection just gained a security dimension. In his Oct 10 essay “Models as Insider Risks in the Super Intelligence Era”, Microsoft CEO Satya Nadella urged treating every frontier model as potentially compromised — separating model intelligence from permissions, keeping tamper-proof human-readable action logs, running independent audits, and installing a human-controlled “emergency brake” to pause or shut down a model mid-task. For builders choosing APIs, this reframes the question: cost per token is only half the equation — governance, auditability and contained deployment (strengths of Alibaba Cloud Model Studio / Bailian) now decide which models enterprises trust in production.
Prices in USD per million tokens, as of August 26, 2026. DeepSeek now uses peak/off-peak pricing (peak: weekdays 9:00-12:00, 14:00-18:00 Beijing time).
| Model | Input | Output | Total (1M+1M) | vs Qwen3.5-Flash |
|---|---|---|---|---|
| Qwen3.5-Flash 🏆 | $0.07 | $0.26 | $0.33 | 1x |
| DeepSeek V4 Flash (off-peak) | $0.22 | $0.66 | $0.88 | 2.7x |
| DeepSeek V4 Flash (peak) | $0.44 | $1.32 | $1.76 | 5.3x |
| DeepSeek V4 Pro (off-peak) | $0.66 | $1.98 | $2.64 | 8x |
| DeepSeek V4 Pro (peak) | $1.32 | $3.96 | $5.28 | 16x |
| Gemini 2.5 Flash | $0.30 | $2.50 | $2.80 | 8.5x |
| Gemini 3.5 Flash | $1.50 | $9.00 | $10.50 | 32x |
| Claude Sonnet 5 (current) | $2.00 | $10.00 | $12.00 | 36x |
| Claude Sonnet 5 (after Sept) | $3.00 | $15.00 | $18.00 | 55x |
| GPT-5.5 | $5.00 | $30.00 | $35.00 | 106x |
| GPT-5.5 Pro | $30.00 | $180.00 | $210.00 | 636x |
Sources: Official pricing pages, DeepSeek API docs, Morgan Stanley research, GitHub state-of-llm-apis, August 2026.
| Model | Input | Output | Context | Notes |
|---|---|---|---|---|
| Qwen3.5-Flash | ¥0.2 | ¥2 | 1M | Multimodal, cache hit ¥0.02 |
| qwen-turbo | ¥0.3 | ¥0.6 | 131K | Thinking mode output ¥3 |
| Qwen3.5-Plus | ¥0.8 | ¥4.8 | 128K | Enhanced reasoning |
| DeepSeek V4 Flash (off-peak) | ¥1.5 | ¥4.5 | — | Weekends all off-peak |
| DeepSeek V4 Flash (peak) | ¥3 | ¥9 | — | Up 200-350% |
| DeepSeek V4 Pro (off-peak) | ¥4.5 | ¥13.5 | — | Cache hit ¥0.15 |
| DeepSeek V4 Pro (peak) | ¥9 | ¥27 | — | Cache hit up 1100% |
Launched August 24, 2026, Alibaba Cloud's Wan3.0 generates up to 30-second videos natively and is the first to accept documents (doc/xls/ppt/pdf/md) as input.
| Resolution | Price/sec | Promo (Aug 24-Sep 23) | 30-sec video |
|---|---|---|---|
| 480p | $0.05 | $0.035 | $1.05 |
| 720p | $0.10 | $0.07 | $2.10 |
| 1080p | $0.20 | $0.14 | $4.20 |
Compared to Google Veo 3.1 at $0.40/sec, Wan3.0 1080p costs only 50% as much. New users get 50 seconds of free video generation.
On September 15, 2026, Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Live targets scale and cost efficiency, with automatic switching across 97 languages mid-conversation, near-real-time visual grounding and background tool calls; Extended Thinking targets high-complexity multi-step reasoning and tops the Artificial Analysis Speech-to-Speech Quality Index (82.6). Pricing: Gemini 3.8 Live audio input $0.005/min and output $0.018/min, available via the Gemini API and Google AI Studio. As frontier voice and reasoning APIs get more competitive, Qwen/Model Studio APIs stand out on cost-efficiency. Source: Google Official Blog (Sep 15, 2026).
On September 17, 2026, PrismML released Ternary Bonsai 2 27B, built on Alibaba's open-weight Qwen3.8 27B. Using ternary {-1,0,+1} weights with FP16 group-wise scaling (1.76 effective bits per weight), the model fits in 5.9GB - 9x smaller than its full-precision counterpart - while retaining 98.2% of aggregate benchmark performance (83.9 vs 85.4), with a 262K-token context and multimodal input, free under Apache 2.0. It reaches up to 143 tokens/s on an RTX 5090 and is roughly 40% more energy-efficient than a full-precision 8B model. As the Qwen ecosystem expands to local and private deployment, developers can run near-flagship capability at a fraction of the hardware cost; hybrid cloud+edge inference is becoming the mainstream architecture, with elastic cloud GPU and Model Studio / Bailian APIs covering the heavy lifting. Source: PrismML (Sep 17, 2026).
On September 18, 2026, TypeSafe AI - founded by Diogo Almeida, former OpenAI researcher and co-inventor of ChatGPT and RLHF - released Jev (backed by a $40M seed round led by DCVC). Jev is not a traditional LLM: it outputs no text, returning calibrated structured decisions (classifications, scores, booleans) against developer-defined schemas, which mathematically eliminates format hallucination. End-to-end latency drops to 70-500ms (40-200x faster than frontier models), with input at ~$0.042 per million tokens and free output. In Vercel's tests, Jev's classifier ran 5-18x faster with higher accuracy than OpenAI's model; in Bryo AI's email-classification test, Gemini was slightly more accurate but 10-20x more expensive. The signal: AI pricing is shifting from generic per-token generation toward per-decision value - the cost structure of high-frequency automation (classification, routing, QA) is being rewritten, and combining Jev-style models with general models like Qwen/Model Studio will further lower enterprises' total AI costs. Source: TechCrunch, TypeSafe AI (Sep 18, 2026).
On September 18, Fields Medalist and SAIR Foundation co-founder Terence Tao announced the Open Math Model Initiative: a community-driven effort, in partnership with academia and industry, to build open-weight models and open-source tooling for mathematics and the basic sciences. The first phase targets everyday research tasks — understanding arguments, verifying references and formalizing proofs — and explicitly frames "open models + affordable compute" as a shared foundation for mathematical research. Weights and code will be released under open licenses such as Apache 2.0 / MIT / CC BY 4.0, and the foundation is inviting partners to contribute funding, compute, expertise and community building. For developers and enterprises, a broader open-model ecosystem keeps pushing down the cost of capable AI — combined with on-demand cloud GPU and low-cost APIs (e.g., Alibaba Cloud Model Studio / Bailian), the barrier to high-performance models keeps falling. Source: Terence Tao's blog, SAIR Foundation (Sep 18, 2026).
Nvidia AI servers up 15%+, HBM supply shortages, DRAM up 90-95% in Q1. GPU rental rates: H100 from $1.70 to $2.35/GPU/hour.
GitHub's official announcement: "Agentic workflows have dramatically increased compute demands, with some single requests exceeding the cost of an entire plan." GitHub Copilot introduced session limits and 7-day token caps in April 2026.
OpenRouter data: DeepSeek-V4-Flash processed 11.31 trillion tokens in a single week (Aug 3-7). OpenCode platform processed 8 trillion tokens in a single day.
Morgan Stanley's report is titled "Farewell to Price Wars, Hello to Intelligence Wars." Tencent Cloud raised prices twice this year; Zhipu AI three times.
Qwen3.5-Flash supports 1M token context, multimodal input (text/image/video), Function Calling, and structured output. It scores 100/100 for pricing value in the coding category (lmmarketcap) and is 98% cheaper than the category average. It handles the vast majority of use cases.
Sign up for Alibaba Cloud and enter the Model Studio console — tokens are automatically credited, valid for 180 days. No credit card required for China region; international version offers $200 credit with a 30-day refund guarantee.
DeepSeek V4 Pro remains competitive for complex reasoning, but at peak pricing it approaches mid-tier international model costs. Use Qwen3.5-Flash daily and switch to V4 Pro only when you need its unique capabilities — preferably off-peak or on weekends.
Alibaba Cloud International offers $200 free credits (30-day refund guarantee), with Qwen3.5-Flash at $0.07/$0.26 per M tokens across Singapore, Germany, and US nodes.
New users: $200 cloud credit + 70M AI tokens + 80+ products, 30-day refund guarantee
Claim $200 Free Credit → ECS from $4.50/moChina users: Get 70M free tokens on Bailian →