GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer
Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi
Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload
gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi
Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa
Hailuo 2.3-fast has the lowest effective price (about $0.00185/s), Kling 3.0 offers the fullest range up to 4K, and Seedance 2.5 costs most but render
GPT-5.6 Sol lists at $4/M input and $20/M output (a promo running through Nov 21). Eleven sellers here post 12 tiers, the deepest at 0.077 of list — $
GPT-6 Astra is [OI]'s flagship model for advanced analysis, software engineering, and deep research, with a 1.05M-token context window, $10/M official
At effective price, the cheapest text model is gpt-6-luna ($0.0045/M input), the cheapest image model is krea-2-turbo ($0.0038/image), and the cheapes
NVIDIA accepted Jensen Huang's resignation on Nov 11, with CFO Colette Kress as interim CEO; the stock fell 14% intraday. The same day Edge0 open-sour
After the same key leaked, disabling took effect in 4 milliseconds, while rotation left the old key a 24-hour grace window. Five real responses with s
Google shipped its first decision model, Gemini Decision 1: scores up to 1,000 questions at once, $0.10 per M input tokens, output free.
Alibaba Cloud Bailian listed Qwen4-Max-Preview on Nov 11: 1M context, deep thinking on by default, but Token Plan subscription only - no pay-as-you-go
Baidu released Wenxin 5.5 Flash Preview on Nov 11: 280B total activating 14B, 1M context, 200+ tokens/s, limited-time half price.
Google opened Gemini 4 Carbon to all developers on Nov 10, previously defenders only. Entry price is $2 in / $10 out per M tokens, doubling Dec 1. Not
Anthropic released Claude Fable 5.2 on Nov 9: two packaging of the same base model, coding score up from 32.4% to 49.1%, API price down 20%.
On the same 30-word rewrite, six models charged 56x apart; all eight questions got answered correctly, so the extra money bought no extra accuracy. Me
Tencent Cloud announced Nov 9 that Hy4 preview (hy4-preview) stops service at 00:00 Beijing time on Dec 7, with Hy4 as the recommended migration targe
Meituan open-sourced LongCat-2.5 on Nov 9 under MIT: 1.6T parameters, 1M context, trained entirely on domestic compute, weights live on Hugging Face.
A DeepMind agent found two weaknesses in 1024-bit RSA, compressing the same attack's reproduction from 1,380 to 391 core-years, running in just 3.2 da
Thinking Machines released Inkling 2 Small at 594B, hitting 86.3% on hard coding problems, promising full Apache 2.0 weights next week.
OpenAI opened GPT-6 Cyber: 48.5% on ExploitGym, 81 new vulnerabilities found in Firefox, and privilege-escalation programs written for 948 Linux bugs.
A 429 isn't always you going too fast. I reproduced three kinds of 429 with a real key, and broke down 1,458 of them from 30 days of logs: about 70% c
Google released Gemini Omni 1.2 Flash on Nov 5: 720p video down to $0.07/second, 30% cheaper than last gen, up to 15-second clips with multi-turn edit
Tencent Cloud announced on Nov 4 that Step 5 Preview's cache-hit input price doubles on Nov 6: Guangzhou 0.35→0.70, Singapore 0.36→0.72 yuan per M tok
Tencent Hunyuan released the Hy4 GA (770B, first with vision and computer use); the same day Baidu open-sourced Wenxin 5.5 Flash Base.
Deep Principle released and open-sourced the 753B science model Hydrogen-1.0, trained on GLM-5.3, with a vendor-claimed 38.3% science-reasoning score
Tencent's Hy4 GA and Alibaba's Qwen4 are set to collide at the Wuzhen summit: the 2026 World Internet Conference opened Nov 3 under the banner of 'ope
The same 404 can mean a wrong path, a misspelled model name, or your own key's discount threshold filtering it out. This article replays every error f
Using one API key across DeepSeek, Claude, and Gemini: the same prompt cost 12x more depending on the model, one model has 5 name spellings with input
Meituan launched LongCat-2.5 on Nov 2: 1.6T total parameters, native multimodal, official limited-time price ¥2/¥8 unchanged from the previous gen. Th
DeepSeek open-sourced a new Super-agent architecture on the evening of Oct 31 (GitHub prototype, MIT): a 'dispatcher + workers' two-tier design that m
As of 2026-11-01, LONG AI lists 190 models, 180 with quotes. This article puts Claude, GPT, Gemini, DeepSeek, Qwen, GLM, Kimi, and MiniMax on one tabl
Tencent Hunyuan delivered a technical report for its open world model Hy-Nomas on Oct 30, open-sourcing the Hy-Nomas-Bench benchmark, then recreated a
Mistral released open weights for flagship Mistral Large 4 (codenamed Le Chonk) on Oct 29: ~1T total / ~52B activated, 1M context, free download on Hu
Google released Gemini 4 Carbon on Oct 28, topping the AA intelligence index at 59 and hitting 88.2% on vulnerability detection. Entry pricing matches
OpenAI released the mid-tier GPT-6.1 Terra on Oct 26-27: $1 in / $5 out per M tokens, 50% below the previous gen and 1/10 of the GPT-6 Astra flagship;
Six days later the plaza shifted again: 593 model names grew to 709; 120 old names vanished (90 truly delisted, 30 just respelled). The 3D category sh
BOSS Zhipin's Nanbeige LLM lab open-sourced Nanbeige4.3-33B on Oct 24-25: a 33B dense architecture, 520K context, Apache 2.0, winning 7 of 8 agentic c
Third-party replication of Kimi K3.1 Pro is in: 55 on the Artificial Analysis intelligence index, #4 among 215 models, up 11 from K3's 44, shrinking t
Moonshot released Kimi K3.1 Pro and its Highspeed variant in the early hours of Oct 24: ~3T parameters, 1M context, vendor-claimed Terminal-Bench 4.0
Ant Digital released the decision model LingDT-Decision-1.0 on Oct 22: 4B and 27B versions that don't write text, only make structured judgments like
iFlytek released Spark X2.6 on Oct 22: a 293B-A30B MoE with context expanded to 1M tokens, the first flagship claiming to be trained entirely on domes
Inception released Mercury 3 on Oct 20: diffusion architecture scaled to 40x of Mercury 1, context up to 1M, vendor-claimed 1,068 tokens/s on NVIDIA G
The plaza list changed on Oct 21: 574 model names became 593; 174 old call names disappeared, 141 truly delisted and 33 just moved to a new channel pr
MiniMax released M3.1 Flash on Oct 19: ~3x faster inference, 70% lower agent-task cost, DeepSWE 73.7; the previously anonymous Space Bunny that topped
MiniMax released M3.1 Flash on Oct 19 - native multimodal, 1M context, official pricing of ¥1 in / ¥4 out, 52% cheaper than the previous M3. But our p
Five real turns, logged token by token: 1,368 tokens total, ¥0.00052 at our prices, ¥5.18 for ten thousand runs. Measured data showing the money goes
A no-nonsense Codex CLI guide: install with three commands, point one config file at a local relay, and watch it actually fix a bug and pass the tests
Nine out of ten DeepSeek top-up failures come down to three things: payment risk control, using the wrong portal, or a delayed balance update. Here's