Blog · Findings and weekly digestLatest 2026-11-17

AI Findings · Data Digest

Findings from BenchGecko's own tests and price tracking as they happen, and a weekly digest of new models, price changes and benchmark results. Every number links to its data.

Finding·2026-11-17

OpenAI's GPT-6.2 Sol on BenchGecko's own tests: language cost, with identity and refusals pending. Price: $2/M input and $10/M output (list price).

Finding·2026-11-17

OpenAI's GPT-6.2 Sol Pro on BenchGecko's own tests: language cost, with identity and refusals pending. Price: $2/M input and $10/M output (list price).

Finding·2026-11-17

Alibaba's Qwen4-Max on BenchGecko's own tests: identity, refusals, knowledge cutoff and weekly drift, with language cost pending. Price: $1.425/M input and $4.275/M output at Alibaba, the cheapest of 1 offers.

Finding·2026-11-17

GLM 4.5 Air's list price dropped 35% (input $0.2/M to $0.13/M, output $1.1/M to $0.85/M), confirmed in the next daily snapshot.

Finding·2026-11-17

GLM 5.3's list price dropped 61% (input $0.084/M to $0.0699/M, output $3.6/M to $1.4/M), confirmed in the next daily snapshot.

Finding·2026-11-16

GLM 4.7's list price dropped 33% (input $0.6/M to $0.4/M, output $2.2/M to $1.75/M), confirmed in the next daily snapshot.

Finding·2026-11-16

DeepSeek V4 Flash 0731's list price dropped 40% (input $0.0434/M to $0.06/M, output $0.3/M to $0.18/M), confirmed in the next daily snapshot.

Weekly digest·2026-11-16

Tencent's cheapest Hy model · 224B total and 12B active parameters · 1M token context · $0.070 input and $0.281 output per million tokens.

Weekly digest·2026-11-16

Alibaba's Qwen4 flagship goes pay per token · $1.50 input and $4.50 output per million tokens · 1M token context · 25% below Qwen3.8 Max list price.

Scorecard·2026-11-16

This week's top 5 on BenchGecko's own tests: 1. Qwen3.8 Max Prime (83), 2. Qwen3.8 27B (81), 3. Claude Opus 5.5 (70), 4. Claude Sonnet 5.5 (69), 5. Qwen3.8 Max (0902) (68).

Finding·2026-11-15

DeepSeek V4.1 Flash's list price dropped 83% (input $0.3/M to $0.0507/M, output $1.2/M to $0.2028/M), confirmed in the next daily snapshot.

Finding·2026-11-15

DeepSeek V4.1 Flash 1017's list price dropped 87% (input $0.3/M to $0.038/M, output $1.2/M to $0.5/M), confirmed in the next daily snapshot.

Finding·2026-11-14

MiMo-V2.5-Pro's list price dropped 52% (input $0.48/M to $0.435/M, output $1.8/M to $0.87/M), confirmed in the next daily snapshot.

Finding·2026-11-13

Kimi K2.6's list price dropped 32% (input $0.95/M to $0.65/M, output $4/M to $3.41/M), confirmed in the next daily snapshot.

Finding·2026-11-13

Granite 4.2 8B's list price dropped 40% (input $0.06/M to $0.06/M, output $0.25/M to $0.15/M), confirmed in the next daily snapshot.

Finding·2026-11-12

Google's Gemini 4 Hydrogen on BenchGecko's own tests: identity, refusals and knowledge cutoff, with language cost pending. Price: $0.75/M input and $3.75/M output at Google AI Studio, the cheapest of 7 offers.

Finding·2026-11-12

Gemma 4 26B A4B 's list price dropped 56% (input $0.0675/M to $0.08/M, output $0.225/M to $0.1/M), confirmed in the next daily snapshot.

Finding·2026-11-12

Qwen3 VL 30B A3B Thinking's list price dropped 58% (input $0.2/M to $0.29/M, output $2.4/M to $1/M), confirmed in the next daily snapshot.

Weekly digest·2026-11-11

Alibaba's first Qwen4 model opens as a two week preview · 1M token context and 393,216 output tokens · Token Plan only, no per token price yet.

Weekly digest·2026-11-10

Google made Gemini 4 Carbon generally available on November 10, at Argon's introductory price until November 30 and $4 input · $20 output per 1M tokens after.

Weekly digest·2026-11-10

Meta released Muse Ember on 2026-11-09. BenchGecko lists it at $0.10 input and $0.50 output per M tokens with a 1M token context and an Artificial Analysis Index of 36.

Weekly digest·2026-11-10

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 2026-11-09. BenchGecko lists Fable 5.2 at $8 input and $40 output per M tokens with a 1M token context.

Finding·2026-11-10

Anthropic's Claude Fable 5.2 on BenchGecko's own tests: identity, language cost and weekly drift. Price: $8/M input and $40/M output at Google, the cheapest of 5 offers.

Finding·2026-11-10

Google's Gemini 4 Carbon on BenchGecko's own tests: identity, refusals and language cost. Price: $1/M input and $5/M output at Google AI Studio, the cheapest of 7 offers.

Finding·2026-11-09

DeepSeek V4.1 Flash (batch)'s list price dropped 90% (input $0.112/M to $0.011/M, output $0.336/M to $0.2/M), confirmed in the next daily snapshot.

Finding·2026-11-09

DeepSeek V4.1 Flash 1017's list price dropped 96% (input $0.15/M to $0.006/M, output $0.6/M to $0.5/M), confirmed in the next daily snapshot.

Weekly digest·2026-11-09

Thinking Machines Lab released Inkling 2 Small on 2026-11-05. BenchGecko lists the 594B total / 17B active multimodal MoE at $0.30/$1.00 per M tokens with a 1M token context on Tinker.

Weekly digest·2026-11-09

Ricursive Intelligence released Ripple 65B Preview on 2026-11-04 under Apache 2.0. BenchGecko lists the 65B dense model at $0.20/$0.60 per M tokens with a 1M token context.

Weekly digest·2026-11-09

Moonshot AI published open weights for Kimi K3.1 Pro on Hugging Face under the Kimi K3.1 Pro License. BenchGecko updates the row to open weights · 2.8T total and 104B active parameters · 1,048,576 token context.

Weekly digest·2026-11-02 to 2026-11-08

9 new models, 22 price changes (11 of them drops) and Mirrorcode results for 12 models were added to BenchGecko. Qwen holds 40.3% of tracked mindshare.

Finding·2026-11-08

Nemotron 3 Super's list price dropped 83% (input $0.3/M to $0.075/M, output $2.4/M to $0.4/M), confirmed in the next daily snapshot.

Finding·2026-11-07

MiniMax M2.7's list price dropped 30% (input $0.3/M to $0.21/M, output $1.2/M to $0.84/M), confirmed in the next daily snapshot.

Finding·2026-11-04

Same 22 probes, same settings, one week apart: its identity answer changed (claims:openai → self).

Weekly digest·2026-11-04

Tencent's flagship Hy4 adds image input and computer use · $0.834 input and $2.501 output per million tokens at list price · 1M token context.

Weekly digest·2026-11-02

Meituan's 1.6T parameter multimodal agent model is now tracked on BenchGecko · $0.30 input and $1.20 output per million tokens · 1M token context.

Finding·2026-11-02

Zhipu AI's GLM 5.3 FlashX Legal on BenchGecko's own tests: identity, refusals and knowledge cutoff, with language cost pending. Price: $0.37/M input and $1.25/M output at Z.AI, the cheapest of 1 offers.

Weekly digest·2026-10-29

Mistral released the weights of Mistral Large 4 on 2026-10-29 and cut its API price from $1.36 to $0.70 input and from $4.18 to $2.10 output per 1M tokens.

Weekly digest·2026-10-29

The Nanbeige LLM Lab released Nanbeige4.3-33B on 2026-10-24 under Apache 2.0. BenchGecko lists the dense 33B model with a 524,288 token context and vendor-reported scores.

Weekly digest·2026-10-29

OpenAI's GPT-6.1 Terra is listed on BenchGecko at $1 per million input tokens and $5 per million output tokens with a 1,050,000 token context window.

Weekly digest·2026-10-29

Liquid AI released the LFM3 family on 2026-10-27 with open weights. BenchGecko lists four models, led by LFM3-1.7B with a 262,144 token context.

Weekly digest·2026-10-29

DiffuSpace released Clover-35B-A3B-Preview on 2026-10-28 under Apache 2.0. BenchGecko lists the block diffusion MoE with 35B total and about 3B active parameters.

Weekly digest·2026-10-29

Gemini 4 Carbon arrives two weeks after Argon at the same introductory price of $2 input and $10 output per 1M tokens, with access limited to cyber defenders for now.

Finding·2026-10-29

Same 22 probes, same settings, one week apart: its identity answer changed (self → claims:anthropic, self → claims:openai).

Finding·2026-10-28

OpenAI's GPT-6.1 Terra on BenchGecko's own tests: identity, refusals, knowledge cutoff, language cost and weekly drift. Price: $0.5/M input and $2.5/M output at OpenAI, the cheapest of 7 offers.

Finding·2026-10-28

OpenAI's GPT-6.1 Terra Pro on BenchGecko's own tests: identity, refusals, knowledge cutoff, language cost and weekly drift. Price: $0.5/M input and $2.5/M output at OpenAI, the cheapest of 6 offers.

Finding·2026-10-26

DeepSeek V4.1 Flash's list price dropped 93% (input $0.3/M to $0.02/M, output $1.2/M to $1/M), confirmed in the next daily snapshot.

Weekly digest·2026-10-26

Recursive released Flint on 2026-10-21 under Apache 2.0. BenchGecko lists the 39.8B MoE with about 2.5B active parameters, a 262,144 token context and a free OpenRouter endpoint.

Weekly digest·2026-10-26

iFlytek released Spark X2.6 on 2026-10-22. BenchGecko lists the closed 293B-A30B MoE with a 1M token context and records iFlytek's yuan promo rates of ¥0.9 input and ¥3.6 output per M tokens.

Weekly digest·2026-10-26

MiniMax published open weights for M3.1 Flash under the MiniMax Community License. BenchGecko updates the row with Hugging Face id MiniMaxAI/MiniMax-M3.1-Flash and keeps the $0.15/$0.60 per M token rates.

Weekly digest·2026-10-19 to 2026-10-25

1 new model, 33 price changes (15 of them price drops) and benchmark results for 21 models added to BenchGecko. MiniMax mindshare rose 799 bps.

Finding·2026-10-25

Asked what it is and who made it, MiniMax's MiniMax M3.1 Flash named OpenAI and Anthropic instead of MiniMax in 2 of 6 replies (run 2026-10-25, no system prompt).

Finding·2026-10-24

Moonshot AI's Kimi K3.1 Pro on BenchGecko's own tests: identity, refusals, knowledge cutoff, language cost and weekly drift. Price: $0.2375/M input and $8.4114/M output at Relace, the cheapest of 20 offers.

Finding·2026-10-24

DeepSeek's DeepSeek V4.1 Flash 1017 on BenchGecko's own tests: identity, refusals, knowledge cutoff and language cost. Price: $0.03/M input and $0.25/M output at Particle AI, the cheapest of 20 offers.

Finding·2026-10-21

Same 22 probes, same settings, one week apart: its identity answer changed (claims:anthropic → self).

Finding·2026-10-21

Beam 1.0's list price dropped 70% (input $1/M to $0.3/M, output $3/M to $1/M), confirmed in the next daily snapshot.

Weekly digest·2026-10-21

Ornith published open weights for Ornith 1.5 122B on 2026-10-20 under MIT. BenchGecko lists it with a 262,144 token context and records Ornith's 84.5 on SWE-bench Verified as a vendor claim.

Weekly digest·2026-10-21

humans& released Ampersand v0.1 on 2026-10-20, a collaborator model post-trained from GLM 5.3. BenchGecko lists it at $1.00 input and $5.20 output per M tokens with a 1,048,576 token context.

Weekly digest·2026-10-20

Inception released Mercury 3 on 2026-10-19. List price is $0.20 input and $0.70 output per M tokens, 50% off at launch, with a 1M token context, up from 260,000 for Mercury 2.5.

Weekly digest·2026-10-20

Meta published the Muse Glimmer 30B paper on 2026-10-19. The weights have been on Hugging Face since 2026-08-09 under Apache 2.0. BenchGecko lists a 131,072 token context and hosted rates from $0.30 input and $1.20 output per M tokens.

Weekly digest·2026-10-19

MiniMax released M3.1 Flash on 2026-10-19. BenchGecko lists it at $0.15 input and $0.60 output per M tokens, half the list price of MiniMax M3, with a 1M token context.

Finding·2026-10-19

Qwen3.8 27B's list price dropped 44% (input $0.425/M to $0.24/M, output $2.55/M to $2.2/M), confirmed in the next daily snapshot.

Finding·2026-10-19

GLM 5.3 Flash's list price dropped 68% (input $0.15/M to $0.048/M, output $0.5/M to $0.2/M), confirmed in the next daily snapshot.

Weekly digest·2026-10-19

Elorian released Glimpse on 2026-10-16 as a research preview. BenchGecko lists it at $0.60 input and $3.00 output per M tokens, with text, image and video inputs.

Weekly digest·2026-10-19

DeepSeek released V4.1 Flash 1017 on 2026-10-17 with MIT weights. BenchGecko lists it at $0.30 input and $1.20 output per M tokens and tracks 15 provider endpoints.

Weekly digest·2026-10-12 to 2026-10-18

Four new models were listed, led by Gemini 4 Argon and Step 5 Preview. BenchGecko recorded 26 price changes, 9 of them drops, and added benchmark results for 29 models.

Finding·2026-10-18

DeepSeek V4 Pro 0813's list price dropped 50% (input $1.32/M to $0.66/M, output $3.96/M to $1.98/M), confirmed in the next daily snapshot.

Finding·2026-10-18

DeepSeek V4 Pro's list price dropped 78% (input $0.95526/M to $0.2088/M, output $1.91052/M to $0.4176/M), confirmed in the next daily snapshot.

Finding·2026-10-17

Qwen3.6 27B's list price dropped 38% (input $0.32/M to $0.25/M, output $3.2/M to $2/M), confirmed in the next daily snapshot.

Finding·2026-10-17

Asked what it is and who made it, DeepSeek's DeepSeek V4.1 Flash 1017 named OpenAI instead of DeepSeek in 5 of 6 replies (run 2026-10-17, no system prompt).

Finding·2026-10-17

Google's Gemini 4 Argon on BenchGecko's own tests: identity, refusals, knowledge cutoff and language cost. Price: $1/M input and $5/M output at Google AI Studio, the cheapest of 6 offers.

Finding·2026-10-17

Mistral AI's Mistral Large 4 on BenchGecko's own tests: identity, refusals, knowledge cutoff and weekly drift, with language cost pending. Price: $0.7/M input and $2.1/M output at Mistral, the cheapest of 2 offers.

Finding·2026-10-17

Google's Nano Banana 2.1 on BenchGecko's own tests: identity and knowledge cutoff. Price: $1.5/M input and $7.5/M output at Google, the cheapest of 2 offers.

Weekly digest·2026-10-16

Reflection released Beam 1.0 open weights on 2026-10-16 under Apache 2.0. BenchGecko lists it at $0.25 input and $1.00 output per M tokens with a 1,048,576 token context window.

Finding·2026-10-16

Kimi K3's list price dropped 53% (input $0.95/M to $0.45/M, output $14/M to $15/M), confirmed in the next daily snapshot.

Weekly digest·2026-10-16

MBZUAI Institute of Foundation Models released Jais 3 on 2026-10-15: six open-weight Arabic-centric models under Apache 2.0, up to 375B total and 23B active parameters, with 512K context on five of the six.

Weekly digest·2026-10-16

Meta released open weights for Muse Spark Open on 2026-10-15. BenchGecko lists it at $1.25 input and $4.25 output per M tokens with a 1,048,576 token context window.

Weekly digest·2026-10-15

StepFun published open weights for Step 5 Preview on 2026-10-15. BenchGecko lists it at $1 input and $2.70 output per M tokens with a 1,048,576 token context window.

Weekly digest·2026-10-15

NVIDIA released Nemotron 3.5 Super VL (120B total, 12B active, up to 1M context) and Nemotron 3.5 Bolt 74B-A9B (24GB, 256K context) as open weights on 2026-10-15.

Weekly digest·2026-10-15

Paper Instruments launched Ultramarine 1 Flash on 2026-10-14, a GLM-5.3 Flash post-train for professional work. BenchGecko lists it at $0.15/$0.50 per M tokens.

Weekly digest·2026-10-15

Liquid AI released LFM2.5-24B-A2B on 2026-10-13. BenchGecko lists it at $0.05/$0.70 per M tokens with a 128,000 token context window.

Weekly digest·2026-10-15

OpenBMB released MiniCPM-V 4.7 on 2026-10-12. BenchGecko lists the 35B-A3B variant, an Apache-2.0 multimodal MoE with a 262,144 token position limit.

Weekly digest·2026-10-15

Meta launched Muse Wonder on 2026-10-12 on Meta Model API and Muse Code. BenchGecko lists it at $4 input, $0.20 cached input and $20 output per M tokens.

Weekly digest·2026-10-15

Google made Gemini 4 Argon generally available on 2026-10-13. BenchGecko lists it at $2/$10 per M tokens (introductory) with a 1,048,576 token context window.

Weekly digest·2026-10-15

Mistral AI opened a public preview of Mistral Large 4 on 2026-10-06. BenchGecko lists the model at $0.68/$2.09 per M tokens (sale) with a 1,048,576 token context.

Weekly digest·2026-10-15

Microsoft released Microsoft-Decision-1 on 2026-10-09 for structured decision scoring. BenchGecko lists it at $0.042 per M input tokens, free output, and 32,768 context.

Finding·2026-10-15

Anthropic's Claude Haiku 5.5 on BenchGecko's own tests: identity, refusals, knowledge cutoff, language cost and weekly drift. Price: $0.1/M input and $0.5/M output at Google, the cheapest of 10 offers.

Finding·2026-10-05

Asked what it is and who made it, DeepSeek's DeepSeek V4 Flash 0731 named Google and Anthropic instead of DeepSeek in 2 of 6 replies (run 2026-10-04, no system prompt).

Finding·2026-10-05

Asked what it is and who made it, DeepSeek's DeepSeek V4.1 Flash named OpenAI instead of DeepSeek in 2 of 5 replies (run 2026-10-04, no system prompt).

Finding·2026-10-04

Asked what it is and who made it, Meta's Llama 4 Maverick named Google and OpenAI instead of Meta in 3 of 6 replies (run 2026-10-04, no system prompt).

Finding·2026-10-04

Asked what it is and who made it, DeepSeek's DeepSeek V4 Pro 0813 named OpenAI and Anthropic instead of DeepSeek in 3 of 6 replies (run 2026-10-04, no system prompt).

Weekly digest·2026-09-26 to 2026-10-02

Six new models were listed, from Anthropic and OpenAI. BenchGecko recorded 26 price changes, 12 of them drops, and added benchmark results for 198 models.