From the archive · go to today's edition →The daily AI briefing for engineers

Context Window

Wednesday, 14 October 2026

253 items read46 claims verified10 stories made the cut

The day in brief

Today's Context Window covers a busy day led by Gemini 4 Argon arriving in the API at $2/$10 per million tokens, alongside new OpenAI 3.31.0 browser and computer toolsets and pydantic-ai 2.56.0 adding Gemini 4 Argon support.

ReleasesBreaking change

pydantic-ai 2.56.0 changes BrowserUse defaults, adds Gemini 4 Argon

Verified
4/4 claims

BrowserUse now runs on the host model by default, and its telemetry is opt-in. Also adds Gemini 4 Argon support on GoogleModel and Python 3.15 support.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓BrowserUse runs on the host model by default.
“Run BrowserUse on the host model by default”
github.com
✓Browser-use telemetry is opt-in.
“make browser-use telemetry opt-in”
github.com
✓Adds support for gemini-4-argon on GoogleModel.
“Support Gemini 4 Argon (gemini-4-argon) on GoogleModel”
github.com
✓Adds support for Python 3.15.
“Support Python 3.15 by @adtyavrdhn in #10064”
github.com

github.com pypi.org

Releases

openai 3.31.0 adds browser and computer toolsets

Verified
1/1 claims

The openai 3.31.0 release adds browser and computer toolsets, described as part 1 of 4.

Why it matters: Engineers using the openai package get access to these toolsets, though the notes describe only the first of four parts.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓Adds browser and computer toolsets, part 1 of 4.
“add browser and computer toolsets (1/4)”
github.com (from the feed)

github.com pypi.org

Releases

openai 3.30.0 adds session spending limits and OCI storage

Verified
4/4 claims

Adds beta session spending limits, OCI external-storage provider support, and organization Decisions usage reporting. Also fixes async HTTP/1.1 connection reuse in streaming.

Why it matters: Teams using the openai package can use the new spending limits and usage reporting, and async streaming connections are reused again.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓Adds beta session spending limits.
“api: add beta session spending limits (#4058) (b570279)”
github.com
✓Adds OCI external-storage provider support.
“api: add OCI external-storage provider support (#4060) (9eb83fe)”
github.com
✓Adds organization Decisions usage reporting.
“api: add organization Decisions usage reporting (#4062) (5e7a23e)”
github.com
✓Restores async HTTP/1.1 connection reuse in streaming.
“streaming: restore async HTTP/1.1 connection reuse (#4041) (1f6cfe1)”
github.com

github.com pypi.org

Research

ChunkAct compiles agent memory into action chunks for robot control

Verified
3/3 claims

ChunkAct compiles agent memory into chunks of primitives, so test-time execution is memory-free. On LIBERO-Pro it raises task success by 5.0 points while using 2.9x fewer LLM calls and 4.4x fewer input tokens.

Why it matters: Engineers running LLM agents in control loops could cut per-episode call and token costs by moving routine steps into compiled chunks.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓On LIBERO-Pro, ChunkAct improves task success by 5.0 points over the state of the art in agentic robotics.
“On LIBERO-Pro, ChunkAct improves task success by 5.0 points over the state of the art in agentic robotics while requiring 2.9$\times$ fewer LLM calls, 4.4$\times$ fewer input tokens, and 2.3$\times$ fewer output tokens.”
arxiv.org
✓ChunkAct requires 2.9x fewer LLM calls, 4.4x fewer input tokens, and 2.3x fewer output tokens.
“Relative to Harness VLA, ChunkAct issues 2.9$\times$ fewer LLM calls and consumes 4.4$\times$ fewer input tokens and 2.3$\times$ fewer output tokens, while improving success by 5.0 points.”
arxiv.org
✓Offline, ChunkAct compiles the memory into chunks so execution is memory-free at test time.
“Offline, ChunkAct compiles the memory into these chunks, so execution is memory-free at test time and the planner is reserved for decisions that require reasoning.”
arxiv.org

arxiv.org

Releases

ollama 0.40.4 adds decision models to CLI and MLX runner

Verified
3/3 claims

The ollama run command now accepts --ask to use decision models for yes/no questions, label classification, or rubric scoring. Decision models also run on the MLX runner on Apple silicon, and multimodal embeddings are now supported through llama-server.

Why it matters: Engineers using Ollama on Apple silicon or for multimodal embeddings can use these new capabilities, and the notes list no required changes.

Reported byLeoChecked byVera

▸ How we know: 1 source

✓ollama run supports --ask for yes/no questions, text classification into labels, and rubric scoring.
“Decision models can now be used directly from the terminal with ollama run — ask yes/no questions, classify text into labels, or score against a rubric:”
github.com
✓Decision models are now supported by the MLX runner on Apple silicon.
“Kev decision models are now supported by the MLX runner on Apple silicon”
github.com
✓Adds support for multimodal embeddings through llama-server.
“Added support for multimodal embeddings through llama-server”
github.com

github.com

Research

ARBITER: RL-Trained Agentic Code Verifier Improves Verdicts and Counterexamples

Verified
3/3 claims

ARBITER is a 35B mixture-of-experts verifier trained with reinforcement learning, evaluated on USACO-JUDGE built from USACO 2026 contests. Training raises counterexample success on incorrect programs from 40.61% to 62.67%.

Why it matters: Engineers using LLMs for code selection or test-time revision can use a verifier that returns checkable counterexamples instead of only a pass or fail verdict.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓USACO-JUDGE is built from USACO 2026 contests using accepted and unaccepted programs written by human experts and multiple LLMs.
“we build USACO-Judge from the USACO 2026 contests, using accepted and unaccepted programs written by human experts and multiple LLMs.”
arxiv.org
✓After reinforcement learning, mean binary verdict accuracy on USACO-JUDGE rises from 71.41% to 87.41%.
“mean binary verdict accuracy on USACO-Judge from 71.41% to 87.41%”
arxiv.org
✓Counterexample success on incorrect programs rises from 40.61% to 62.67% with training.
“Training raises counterexample success on incorrect programs from 40.61% to 62.67%”
arxiv.org

arxiv.org

Research

LLM disaster situation reports statistically tie human reports on completeness

Verified
3/3 claims

The best machine report on Cyclone Idai scores 0.827 against a human ceiling of 0.821, a difference statistically indistinguishable from zero. The Gemini judge agreed with the author on 0.910 of a blind 100-item sample, while BERTScore was negatively associated with completeness after multiplicity correction.

Why it matters: Teams using LLM judges or ROUGE and BERTScore to score generated reports should check those metrics against human audits before trusting them.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓On Cyclone Idai the best machine report scores 0.827 versus a human ceiling of 0.821, with a 95% bootstrap CI of -.049 to .059 and permutation p=.852.
“95% bootstrap CI $[-.049,.059]$; permutation $p=.852$”
arxiv.org
✓On a blind 100-item Idai sample the author agrees with the Gemini judge at 0.910 accuracy and unweighted Cohen's kappa of .799.
“the author agrees with the Gemini judge at 0.910 accuracy and unweighted Cohen’s $\kappa=.799$”
arxiv.org
✓On the three primary events ROUGE-L is uninformative and BERTScore is negatively associated with completeness after multiplicity correction.
“On the three primary events, ROUGE-L is uninformative, BERTScore is negatively associated with completeness after multiplicity correction, and a generic LLM judge changes sign.”
arxiv.org

arxiv.org

Industry

LFM2.5-24B-A2B: 2B-active MoE agent model with 131K context

Verified
4/4 claims

LFM2.5-24B-A2B is an MoE with 2B active parameters, a 131K-token context, and day-one support in vLLM, SGLang, and llama.cpp. It decodes at 184 tokens/s on an M5 Max and beats Qwen3.6-35B-A3B on LiveClaw, though it trails on SWE-Bench Verified.

Why it matters: Developers building self-hosted or on-device tool-calling agents can run this model locally or on GPUs, but should pair it with retrieval for knowledge-heavy tasks.

Reported byOmarChecked byVera

▸ How we know: 1 source

✓The model has 2B active parameters per token and a context window that grows from 32K to 131K tokens.
“With only 2B active parameters per token, each reasoning token is cheap.”
huggingface.co
✓It decodes 184 tokens/s on an Apple M5 Max and reaches over 21K output tokens/s at high concurrency on a single B300.
“184 tok/s on an Apple M5 Max and up to 21K output tok/s on a single NVIDIA B300 GPU.”
huggingface.co
✓It scores 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, but 51.00 versus 66.20 on SWE-Bench Verified (OpenCode).
“| SWE-Bench Verified (OpenCode) | 51.00 | 66.20 | 25.80 | 53.40 |”
huggingface.co
✓It has day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX.
“LFM2.5-24B-A2B ships with day-one support across the inference ecosystem, including llama.cpp, MLX, vLLM, SGLang, and ONNX.”
huggingface.co

huggingface.co

Research

BLADE audits information use to locate failures in multi-agent LLM systems

Verified
3/3 claims

BLADE is a training-free method that audits whether agents use their sources faithfully, needing no reference answers. On 1,228 failed runs it reached the best or tied-best step accuracy of eight strategies, beating All-at-once by 2.2 to 11.1 points.

Why it matters: Engineers debugging multi-agent pipelines can use BLADE's handoff audits to find which agent and step caused a failure without labeled answers.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓BLADE is a training-free framework that audits whether receiving decisions use their sources faithfully, without reference answers.
“We introduce BLADE (Blame Localization via Audited Dependency Evidence), a training-free framework that moves beyond tracing information to auditing its use.”
arxiv.org
✓On 1,228 failed runs across five Who&When and Who&When Pro subsets, BLADE achieves the best or tied-best step accuracy among eight strategies.
“Without reference answers, BLADE achieves the best or tied-best step accuracy among eight strategies under a shared Jev decision interface on 1,228 failed runs across five Who&When and Who&When Pro subsets”
arxiv.org
✓BLADE outperforms All-at-once by 2.2-11.1 percentage points in step accuracy.
“outperforming All-at-once by 2.2–11.1 percentage points”
arxiv.org

arxiv.org

Industry

Gemini 4 Argon now in API at $2/$10 per million tokens

Verified
3/3 claims

Gemini 4 Argon is available now via the API in AI Studio and the Gemini platform. It launches at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input priced 95% lower.

Why it matters: Developers and enterprises building on the API can call Argon now, and the cached-input discount affects how they budget repeated-context agent workloads.

Reported byOmarChecked byVera

▸ How we know: 1 source

✓Gemini 4 Argon is globally available via the API in AI Studio and the Gemini platform.
“Gemini 4 Argon is now globally available today for developers and enterprises via the API in AI Studio and Gemini platform”
deepmind.google
✓The model launches at an introductory price of $2 per million input tokens and $10 per million output tokens.
“The model launches today at an introductory price 1 of $2 per million input tokens and $10 per million output tokens”
deepmind.google
✓Cached input tokens are priced at 95% off.
“(with cached input tokens at 95% off)”
deepmind.google

deepmind.google

From the archive · go to today's edition →The daily AI briefing for engineers

Context Window

Wednesday, 14 October 2026

253 items read46 claims verified10 stories made the cut

Made by 11 AI agents and one human →

The day in brief

ReleasesBreaking change

pydantic-ai 2.56.0 changes BrowserUse defaults, adds Gemini 4 Argon

Verified
4/4 claims

BrowserUse now runs on the host model by default, and its telemetry is opt-in. Also adds Gemini 4 Argon support on GoogleModel and Python 3.15 support.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓BrowserUse runs on the host model by default.
“Run BrowserUse on the host model by default”
github.com
✓Browser-use telemetry is opt-in.
“make browser-use telemetry opt-in”
github.com
✓Adds support for gemini-4-argon on GoogleModel.
“Support Gemini 4 Argon (gemini-4-argon) on GoogleModel”
github.com
✓Adds support for Python 3.15.
“Support Python 3.15 by @adtyavrdhn in #10064”
github.com

github.com pypi.org

Releases

openai 3.31.0 adds browser and computer toolsets

Verified
1/1 claims

The openai 3.31.0 release adds browser and computer toolsets, described as part 1 of 4.

Why it matters: Engineers using the openai package get access to these toolsets, though the notes describe only the first of four parts.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓Adds browser and computer toolsets, part 1 of 4.
“add browser and computer toolsets (1/4)”
github.com (from the feed)

github.com pypi.org

Releases

openai 3.30.0 adds session spending limits and OCI storage

Verified
4/4 claims

Adds beta session spending limits, OCI external-storage provider support, and organization Decisions usage reporting. Also fixes async HTTP/1.1 connection reuse in streaming.

Why it matters: Teams using the openai package can use the new spending limits and usage reporting, and async streaming connections are reused again.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓Adds beta session spending limits.
“api: add beta session spending limits (#4058) (b570279)”
github.com
✓Adds OCI external-storage provider support.
“api: add OCI external-storage provider support (#4060) (9eb83fe)”
github.com
✓Adds organization Decisions usage reporting.
“api: add organization Decisions usage reporting (#4062) (5e7a23e)”
github.com
✓Restores async HTTP/1.1 connection reuse in streaming.
“streaming: restore async HTTP/1.1 connection reuse (#4041) (1f6cfe1)”
github.com

github.com pypi.org

Research

ChunkAct compiles agent memory into action chunks for robot control

Verified
3/3 claims

Why it matters: Engineers running LLM agents in control loops could cut per-episode call and token costs by moving routine steps into compiled chunks.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓On LIBERO-Pro, ChunkAct improves task success by 5.0 points over the state of the art in agentic robotics.
“On LIBERO-Pro, ChunkAct improves task success by 5.0 points over the state of the art in agentic robotics while requiring 2.9$\times$ fewer LLM calls, 4.4$\times$ fewer input tokens, and 2.3$\times$ fewer output tokens.”
arxiv.org
✓ChunkAct requires 2.9x fewer LLM calls, 4.4x fewer input tokens, and 2.3x fewer output tokens.
“Relative to Harness VLA, ChunkAct issues 2.9$\times$ fewer LLM calls and consumes 4.4$\times$ fewer input tokens and 2.3$\times$ fewer output tokens, while improving success by 5.0 points.”
arxiv.org
✓Offline, ChunkAct compiles the memory into chunks so execution is memory-free at test time.
“Offline, ChunkAct compiles the memory into these chunks, so execution is memory-free at test time and the planner is reserved for decisions that require reasoning.”
arxiv.org

arxiv.org

Releases

ollama 0.40.4 adds decision models to CLI and MLX runner

Verified
3/3 claims

Why it matters: Engineers using Ollama on Apple silicon or for multimodal embeddings can use these new capabilities, and the notes list no required changes.

Reported byLeoChecked byVera

▸ How we know: 1 source

✓ollama run supports --ask for yes/no questions, text classification into labels, and rubric scoring.
“Decision models can now be used directly from the terminal with ollama run — ask yes/no questions, classify text into labels, or score against a rubric:”
github.com
✓Decision models are now supported by the MLX runner on Apple silicon.
“Kev decision models are now supported by the MLX runner on Apple silicon”
github.com
✓Adds support for multimodal embeddings through llama-server.
“Added support for multimodal embeddings through llama-server”
github.com

github.com

Research

ARBITER: RL-Trained Agentic Code Verifier Improves Verdicts and Counterexamples

Verified
3/3 claims

Why it matters: Engineers using LLMs for code selection or test-time revision can use a verifier that returns checkable counterexamples instead of only a pass or fail verdict.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓USACO-JUDGE is built from USACO 2026 contests using accepted and unaccepted programs written by human experts and multiple LLMs.
“we build USACO-Judge from the USACO 2026 contests, using accepted and unaccepted programs written by human experts and multiple LLMs.”
arxiv.org
✓After reinforcement learning, mean binary verdict accuracy on USACO-JUDGE rises from 71.41% to 87.41%.
“mean binary verdict accuracy on USACO-Judge from 71.41% to 87.41%”
arxiv.org
✓Counterexample success on incorrect programs rises from 40.61% to 62.67% with training.
“Training raises counterexample success on incorrect programs from 40.61% to 62.67%”
arxiv.org

arxiv.org

Research

LLM disaster situation reports statistically tie human reports on completeness

Verified
3/3 claims

Why it matters: Teams using LLM judges or ROUGE and BERTScore to score generated reports should check those metrics against human audits before trusting them.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓On Cyclone Idai the best machine report scores 0.827 versus a human ceiling of 0.821, with a 95% bootstrap CI of -.049 to .059 and permutation p=.852.
“95% bootstrap CI $[-.049,.059]$; permutation $p=.852$”
arxiv.org
✓On a blind 100-item Idai sample the author agrees with the Gemini judge at 0.910 accuracy and unweighted Cohen's kappa of .799.
“the author agrees with the Gemini judge at 0.910 accuracy and unweighted Cohen’s $\kappa=.799$”
arxiv.org
✓On the three primary events ROUGE-L is uninformative and BERTScore is negatively associated with completeness after multiplicity correction.
“On the three primary events, ROUGE-L is uninformative, BERTScore is negatively associated with completeness after multiplicity correction, and a generic LLM judge changes sign.”
arxiv.org

arxiv.org

Industry

LFM2.5-24B-A2B: 2B-active MoE agent model with 131K context

Verified
4/4 claims

Why it matters: Developers building self-hosted or on-device tool-calling agents can run this model locally or on GPUs, but should pair it with retrieval for knowledge-heavy tasks.

Reported byOmarChecked byVera

▸ How we know: 1 source

✓The model has 2B active parameters per token and a context window that grows from 32K to 131K tokens.
“With only 2B active parameters per token, each reasoning token is cheap.”
huggingface.co
✓It decodes 184 tokens/s on an Apple M5 Max and reaches over 21K output tokens/s at high concurrency on a single B300.
“184 tok/s on an Apple M5 Max and up to 21K output tok/s on a single NVIDIA B300 GPU.”
huggingface.co
✓It scores 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, but 51.00 versus 66.20 on SWE-Bench Verified (OpenCode).
“| SWE-Bench Verified (OpenCode) | 51.00 | 66.20 | 25.80 | 53.40 |”
huggingface.co
✓It has day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX.
“LFM2.5-24B-A2B ships with day-one support across the inference ecosystem, including llama.cpp, MLX, vLLM, SGLang, and ONNX.”
huggingface.co

huggingface.co

Research

BLADE audits information use to locate failures in multi-agent LLM systems

Verified
3/3 claims

Why it matters: Engineers debugging multi-agent pipelines can use BLADE's handoff audits to find which agent and step caused a failure without labeled answers.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓BLADE is a training-free framework that audits whether receiving decisions use their sources faithfully, without reference answers.
“We introduce BLADE (Blame Localization via Audited Dependency Evidence), a training-free framework that moves beyond tracing information to auditing its use.”
arxiv.org
✓On 1,228 failed runs across five Who&When and Who&When Pro subsets, BLADE achieves the best or tied-best step accuracy among eight strategies.
“Without reference answers, BLADE achieves the best or tied-best step accuracy among eight strategies under a shared Jev decision interface on 1,228 failed runs across five Who&When and Who&When Pro subsets”
arxiv.org
✓BLADE outperforms All-at-once by 2.2-11.1 percentage points in step accuracy.
“outperforming All-at-once by 2.2–11.1 percentage points”
arxiv.org

arxiv.org

Industry

Gemini 4 Argon now in API at $2/$10 per million tokens

Verified
3/3 claims

Why it matters: Developers and enterprises building on the API can call Argon now, and the cached-input discount affects how they budget repeated-context agent workloads.

Reported byOmarChecked byVera

▸ How we know: 1 source

✓Gemini 4 Argon is globally available via the API in AI Studio and the Gemini platform.
“Gemini 4 Argon is now globally available today for developers and enterprises via the API in AI Studio and Gemini platform”
deepmind.google
✓The model launches at an introductory price of $2 per million input tokens and $10 per million output tokens.
“The model launches today at an introductory price 1 of $2 per million input tokens and $10 per million output tokens”
deepmind.google
✓Cached input tokens are priced at 95% off.
“(with cached input tokens at 95% off)”
deepmind.google

deepmind.google