Contents factory ledger. Files on GitHub, hub on Cloudflare, branches to other platforms.
Lab Notes scoreboard
2026-09-02 01:01 · slot=us-lunch lane=score draft=google/gemini-2.5-flash voice=lab-en 2026-09-02 01:01 · 밖→안(팩토리 加工)→밖 · 다소스intake+순서강건교차검증+링크게이트 · 게시 前 사람 승인 intake 소스 2: perplexity, perplexity_cross · 링크 신뢰1·미지13·죽음1(도달성≠신뢰성)
Lab Notes: August 2026 Price and Performance Benchmarks
This public ledger documents recent observations regarding model pricing and performance metrics, gathered from various sources.
Key Observations
Gemini Deep Research Agent pricing is confirmed. The listed rates are $2 per 1 million input tokens, $0.2 per 1 million cached input tokens, and $12 for output. Web Search Request pricing, for Claude-supported search agents on the same platform, is listed at $10 per 1,000 searches. Grok 4.6 pricing shows tiered rates based on context length. These are $2 per 1 million tokens for input and $6 per 1 million tokens for output for contexts under 200K, and $4 per 1 million tokens for input and $12 per 1 million tokens for output for contexts above 200K.
Grok 4 API pricing appears in a separate August 2026 roundup. The listed rates are $5 per 1 million input tokens and $15 per 1 million output tokens. Grok 4 mini is listed at $0.30 per 1 million input tokens and $0.50 per 1 million output tokens. Gemini 3.1 Pro preview pricing is noted in a later August 2026 cost report. This lists $2 per 1 million input tokens and $12 per 1 million output tokens.
Gemini 3.7 Flash is reported with performance figures. It achieved 340.1 tokens per second and an index score of 56. Qwen3.8-Flash-Next showed the lowest time-to-first-answer-token in provider benchmarks, recorded at 25.44 seconds. Qwen3.8-27B provider benchmarking included DeepInfra timing and score data: $0.31 for an unnamed metric, 47.30 seconds, and a score of 61.51.
Qwen 3.8 Max was launched as a 2.4T Mixture-of-Experts (MoE) model with 95 billion active parameters and a 1 million token context window. Its pricing is given as $2 per 1 million tokens for input and $6 per 1 million tokens for output. Benchmark claims for Qwen 3.8 Max include scores of 93.0 on PaperBench, 67.7 on SWE-bench Pro, and 56.6 on DeepSWE.
DeepSeek V4-Pro-0813 is described as a 1.6T MoE model with API pricing. Off-peak rates are $0.66 per 1 million input tokens and $1.98 per 1 million output tokens. Peak rates are $1.32 per 1 million input tokens and $3.96 per 1 million output tokens. DeepSeek V4-Pro-0813 reportedly achieved 96.40% performance on SWE-bench Verified.
GLM-5.3 was presented as an open-weight coding model with a delayed weight release, approximately two weeks after launch. GLM-5.3 was reported to deliver a 6x gain on Terminal-Bench. Huawei open-sourced openPangu-2.0-Pro as a 505 billion-parameter open-weight model.
Qwen 3.8 and other open models were flagged in a China AI weekly roundup. They were expected by the end of August.
Data Table
| claim | label | order (🟢 robust / ⚠️ sensitive) | numbers | URL |
|---|---|---|---|---|
| Gemini Deep Research Agent pricing is listed on Google’s agent platform page | 🟢 confirmed | 🟢 robust | $2 / 1M input; $0.2 / 1M cached input; $12 output | https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing |
| Web Search Request pricing on the same page is listed for Claude-supported search agents | 🟢 confirmed | 🟢 robust | $10 / 1,000 searches | https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing |
| Grok 4.6 pricing is listed with tiered rates by context length | 🟢 confirmed | 🟢 robust | $2/$6 under 200K; $4/$12 above 200K | https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing |
| Grok 4 API pricing is listed in a separate August 2026 pricing roundup | 🟢 confirmed | ⚠️ sensitive | $5/$15; Grok 4 mini $0.30/$0.50 | https://www.swfte.com/api-pricing |
| Gemini 3.1 Pro preview pricing is listed in a later August 2026 cost report | 🟡 partial | ⚠️ sensitive | $2 / 1M input; $12 / 1M output | https://cutoffreport.com/posts/cutoff-data-brief-2026-08-07 |
| Gemini 3.7 Flash is reported with strong speed and index scores | 🟡 partial | ⚠️ sensitive | 340.1 tokens/sec; index 56 | https://www.buildfastwithai.com/blogs/ai-news-today-august-22-23-2026 |
| Qwen3.8-Flash-Next had the lowest time-to-first-answer-token in provider benchmarks | 🟡 partial | ⚠️ sensitive | 25.44 s | https://artificialanalysis.ai/models/qwen3-8-flash-next/providers |
| Qwen3.8-27B provider benchmarking showed DeepInfra timing and score data | 🟡 partial | ⚠️ sensitive | $0.31; 47.30 s; 61.51 score | https://artificialanalysis.ai/models/qwen3-8-27b/providers |
| Qwen 3.8 Max was launched as a 2.4T MoE with 95B active parameters and 1M context | 🟡 partial | ⚠️ sensitive | 2.4T; 95B active; 1M context; $2/$6 per 1M tokens | https://www.developersdigest.tech/blog/qwen-3-8-max-release-2026 |
| Qwen 3.8 Max benchmark claims include PaperBench, SWE-bench Pro, and DeepSWE | 🟡 partial | ⚠️ sensitive | 93.0; 67.7; 56.6 | https://www.developersdigest.tech/blog/qwen-3-8-max-release-2026 |
| DeepSeek V4-Pro-0813 is described as a 1.6T MoE model with API pricing | 🟡 partial | ⚠️ sensitive | 1.6T; $0.66/$1.98 off-peak; $1.32/$3.96 peak | https://flowtivity.ai/blog/glm-5-3-vs-deepseek-v4-pro-comparison/ |
| DeepSeek V4-Pro-0813 reportedly reached strong SWE-bench Verified performance | 🟡 partial | ⚠️ sensitive | 96.40% | https://flowtivity.ai/blog/glm-5-3-vs-deepseek-v4-pro-comparison/ |
| GLM-5.3 was presented as an open-weight coding model with delayed weight release | 🟡 partial | ⚠️ sensitive | weights promised ~2 weeks after launch | https://apidog.com/blog/what-is-glm-5-3/ |
| GLM-5.3 was reported to deliver a large Terminal-Bench gain | 🟡 partial | ⚠️ sensitive | 6x Terminal-Bench gain | https://zenn.dev/neotechpark/articles/7003e20300d7e5 |
| Huawei open-sourced openPangu-2.0-Pro as a 505B-parameter open-weight model | 🟡 partial | ⚠️ sensitive | 505B | https://www.opensourceforu.com/2026/08/huawei-open-sources-505b-openpangu-ai-drops-weights-and-code/ |
| Qwen 3.8 and other open models were flagged in a China AI weekly roundup | 🔴 hype | ⚠️ sensitive | expected by end of August | https://www.bighatgroup.com/blog/china-ai-weekly-2026-08-01/ |
Practitioner Lines
- Act: Review the confirmed agent pricing structures for Gemini and Claude-supported search agents to optimize operational costs for confirmed use cases.
- Watch: Monitor the actual release and performance metrics of Grok 4 API and Gemini 3.1 Pro once their sensitive pricing and performance data are fully confirmed and made robust.
- Ignore: Dismiss the general expectation of “Qwen 3.8 and other open models” by a specific date, as it lacks precise detail and is categorized as hype.
⚠️ 加工엔진(멀티모델 검증)이 主이지 원료(밖)나 저자(안)가 主 아님. ①출처실재·②순서강건 교차검증만 자동. 최종 게시·맥락판단은 사람.