Every figure this video puts on screen, and where it came from. Checked 31 July 2026.
Laguna S 2.1 was released by poolside on 21 July 2026. Source: poolside, “Introducing Laguna S 2.1”, https://poolside.ai/blog/introducing-laguna-s-2-1
118 billion total parameters, roughly 8 billion activated per token. It is a mixture of experts model built for agentic software engineering. Source: poolside launch post, and the Hugging Face model card at https://huggingface.co/poolside/Laguna-S-2.1
The weights are published on Hugging Face under the OpenMDW-1.1 licence, which permits commercial and non-commercial use and modification. BF16, FP8, INT4 and NVFP4 checkpoints are published, along with GGUF and MLX conversions. Sources: https://huggingface.co/poolside/Laguna-S-2.1, https://huggingface.co/poolside/Laguna-S-2.1-FP8, https://huggingface.co/poolside/Laguna-S-2.1-NVFP4
Context: 1,048,576 tokens on the paid endpoint, 262,144 on the free one. The published checkpoint is configured for 262,144 tokens; training included a long context extension stage to 1,048,576, and poolside states 1M in both thinking and no thinking modes. Sources: poolside launch post; Hugging Face model card; https://openrouter.ai/poolside/laguna-s-2.1 and https://openrouter.ai/poolside/laguna-s-2.1:free
Terminal-Bench 2.1: 70.2% with max thinking, 60.4% with thinking off. Terminal-Bench runs long horizon tasks in a real shell and grades the end state of the machine rather than an answer. Source: poolside launch post, benchmark table
DeepSeek V4 Pro Max scores 64.0% on the same table, at 1.6 trillion total parameters with 49 billion active. That is the 6.2 point gap and the roughly 13x parameter difference the video states (1.6T / 118B = 13.6). Source: poolside launch post, benchmark table
One open model on that table scores higher: Tencent Hy3, 295B total with 21B active, at 71.7%. Laguna S 2.1 is not the top open model on this benchmark. Source: poolside launch post, benchmark table
Other results poolside published on the same day, not used on screen but relevant to how broad the claim is: SWE-Bench Multilingual 78.5% thinking / 71% no thinking, SWE-Bench Pro (public) 59.4% / 53%, DeepSWE v1.1 40.4% / 16.5%, SWE Atlas 46.2%, Toolathlon Verified 49.7%. Source: poolside launch post
Caveat carried on screen: these are the vendor’s own published figures. The video says so in the shot rather than in a footnote, because independent reruns of a ten day old position are thin.
Laguna S 2.1 on OpenRouter lists at $0.10 per million input tokens and $0.20 per million output tokens, with cache reads at $0.01 per million. At the time of checking the model page showed a 10% discount, $0.09 and $0.18, which is why the video says the listed price moves. Sources: poolside launch post (list price); https://openrouter.ai/poolside/laguna-s-2.1 (live card, 31 July 2026)
The free endpoint is free, at 262,144 tokens of context, and OpenRouter states that inputs and outputs on the free route may be used to train and improve models. Source: https://openrouter.ai/poolside/laguna-s-2.1:free
One provider serves the model on OpenRouter. The model page states that OpenRouter forwards every request to a single host with no routing decision to make, so there is no failover and no second price. Source: https://openrouter.ai/poolside/laguna-s-2.1, 31 July 2026
GPT-5.6 Terra is $2.00 per million input tokens and $12.00 per million output tokens, after OpenAI cut Terra by 20% and Luna by 80% on 30 July 2026. Sol was unchanged at $5.00 / $30.00 and Luna is $0.20 / $1.20. Sources: https://www.aipricing.guru/openai-pricing/ (read 31 July 2026); OpenAI, “Advancing the price-performance frontier with GPT-5.6”, https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/
Everything below is worked from the list prices above. It is a modelled example, not a measurement of anyone’s invoice, and the video says so on screen.
One session, assumed at 2,000,000 input tokens and 250,000 output tokens:
One developer, twenty sessions a day, twenty two working days: 440 sessions.
The session size is an assumption. The prices it multiplies are not.
Weights at each precision, from the parameter count:
These are the weights alone, before the KV cache, the activations or any serving overhead.
poolside states the model is small enough to run on a single NVIDIA DGX Spark, which carries 128 GB of unified memory shared between CPU and GPU. The four bit checkpoint at roughly 59 GB is what makes that fit with room left for the cache; the FP8 checkpoint at 118 GB does not leave meaningful headroom on a 128 GB machine. Sources: poolside launch post; NVIDIA DGX Spark product page, https://www.nvidia.com/en-us/products/workstations/dgx-spark/ and the DGX Spark system overview, https://docs.nvidia.com/dgx/dgx-spark/system-overview.html