INSIGHT AI

Source-linked text edition ·

AI Daily

The written, source-linked counterpart to this video.

Published edition11 stories

Bedrock searches video scenes with Marengo 3.0

AWS added TwelveLabs Marengo 3.0 multimodal embeddings to Bedrock Managed Knowledge Base.

Results include segment start and end times, enabling applications to jump directly to the relevant moment in a video.

The announcement does not establish error-free retrieval or specify regional eligibility and pricing.

Watch next
Compare Marengo retrieval accuracy in a benchmark of known video moments and their returned timestamps.
Source evidence for Bedrock searches video scenes with Marengo 3.0
Source-linked proof · original source

Alibaba Token Plan adds metered agent tools

Alibaba says its personal Token Plan added Agent Harness tool benefits effective September 11 without changing existing prices or Credit allowances.

原有价格及Credit额度保持不变, Standard和Pro套餐新增Agent Harness工具权益与用量, 覆盖搜索、网页解析、图像生成、语音处理、代码执行等12项Agent开发常用能力。.

Standard and Pro subscriptions add allowances for search, webpage parsing, code execution, speech and image tools through MCP. Subscribe on the Qianwen AI platform, obtain the Token Plan endpoint and API key, then configure a compatible tool. Included allowances vary by tool and plan. After they run out, additional calls are billed; video generation and several managed services are pay-as-you-go rather than included unlimited usage.

Watch next
Check the Standard or Pro tool quota before a workflow; monthly tool allowances are separate from model Credits.
Source evidence for Alibaba Token Plan adds metered agent tools
Source-linked proof · original source

Claude plugin eval compares with and without your plugin

Claude Code added plugin behavior evaluation with a no-plugin baseline.

Use evals to measure how reliably your plugin steers Claude to the right outcome, to catch regressions when you change the plugin or a new model ships, and to see what the plugin contributes compared with no plugin at all.

Plugin authors need Claude Code 2.1.269 or later and a plugin manifest. Run claude plugin eval init at the plugin root, then claude plugin eval dot to test prompts against graders. By default each case runs three times with the plugin and three without. These and judge calls consume plan usage or API billing. A high score alone does not prove the plugin helped; compare the two arms.

Watch next
Inspect failed graders and usage-limit errors before interpreting a score change as a regression.
Source evidence for Claude plugin eval compares with and without your plugin
Source-linked proof · original source

Cognition brings SWE-2 to Devin Desktop and CLI

Cognition released SWE-2, post-trained from Kimi K3, and made it available in Devin Desktop and CLI.

SWE-2 is available starting today in Devin Desktop and CLI. We’re also rolling it out on Devin Web and Fusion.

Engineers can use SWE-2 in those clients for coding work; Devin Web and Fusion are still described as rolling out. Cognition reports fewer detours and earlier implementation in its task benchmarks. Those are vendor evaluations under specific effort levels and task sets. Its pricing and benchmark comparisons are not guarantees that an individual project will be cheaper or correctly completed.

Watch next
Compare accepted code and total cost in a SWE-2 benchmark using the same project and test suite.
Source evidence for Cognition brings SWE-2 to Devin Desktop and CLI
Source-linked proof · original source

Kimi K2.8 Preview keeps the existing coding model ID

Kimi's September 11 update rolls K2.8 Preview into Kimi Code while keeping the kimi-for-coding model ID.

K2.8 Preview 现已全量上线 Kimi Code, kimi-for-coding 直接升级、无需修改配置: 综合性能接近 K3, 支持 low / high / max 三档思考程度与最高 1M 上下文。详见 最新动态 。.

The claim that performance approaches K3 is Kimi's assessment, not our independent benchmark.

Watch next
Compare task quality and token cost in a K2.8 Preview benchmark that keeps reasoning effort constant.
Source evidence for Kimi K2.8 Preview keeps the existing coding model ID
Source-linked proof · original source

Sakana separates cost-focused Fugu Max from Ultra v2

Sakana released Fugu Max and Fugu Ultra v2, two configurations of its model-orchestration architecture.

Both models are available today via our standard OpenAI-compatible API.

The benchmark comparisons are Sakana's claims, and SWEFish is its internal coding benchmark. They do not establish a guaranteed advantage on a customer's workload.

Watch next
Compare Max and Ultra v2 on a workload benchmark using accepted outputs and total task cost.
Source evidence for Sakana separates cost-focused Fugu Max from Ultra v2
Source-linked proof · original source

AWS compares model cost per successful outcome

AWS published an open benchmarking harness comparing practical model deployments by cost per correct answer and completed task.

This is a comparison of practical deployment configurations, not a controlled estimate of intrinsic model capability.

The results do not isolate intrinsic model capability.

Watch next
Compare model cost per accepted deliverable in a benchmark that includes failed attempts and review work.
Source evidence for AWS compares model cost per successful outcome
Source-linked proof · original source

AgentCore separates working infrastructure from successful tasks

AWS published a dual-monitoring reference for an airline-reservation agent system.

Infrastructure monitoring and agent effectiveness monitoring require different approaches.

AgentCore Evaluations scores sampled interactions for helpfulness, correctness and goal completion. AWS DevOps Agent investigates infrastructure failures such as a missing model-invocation permission. Setup requires AWS account permissions and the Boto3 SDK. Model judges can be wrong, and asynchronous sampling can miss or detect a bad response only after delivery. This is an architecture example, not guaranteed autonomous incident resolution.

Watch next
Compare evaluation judgments against known failures in an AgentCore benchmark; inspect missed goals despite healthy infrastructure.
Source evidence for AgentCore separates working infrastructure from successful tasks
Source-linked proof · original source

AgentCore turns MCP responses into interactive booking cards

AWS published an AgentCore MCP Apps example that returns interactive HTML widgets inside compatible AI hosts.

If the tool has an associated resource URI (for example, ui: //widget/unicorn-list ), the AI host initiates this phase. Tools without an associated widget, such as view_bookings and return_unicorn, return text-only content and skip this phase entirely.

The sample uses an unauthenticated Gateway fronted by WAF; production access controls still need deliberate design.

Watch next
Check the next AgentCore MCP Apps release note for rollout, availability, or quota changes.
Source evidence for AgentCore turns MCP responses into interactive booking cards
Source-linked proof · original source

NVIDIA NIM benchmark depends on workload and hardware

NVIDIA published a Nemotron 3 Ultra NIM serving configuration and performance comparison.

The published curves are a starting point, not a promise that every application will see the same result.

The reported throughput uplift combines caching, scheduling, precision and speculative decoding. It is not a general model-quality improvement, and the component gains cannot simply be added together.

Watch next
Compare NIM configurations on a benchmark with the same hardware and per-user latency target.
Source evidence for NVIDIA NIM benchmark depends on workload and hardware
Source-linked proof · original source

SageMaker routes shared prompts to warm caches

AWS added prefix-aware routing for SageMaker real-time inference endpoints.

You need at least two instances. With one instance, all requests go to the same place regardless of strategy.

The existing Invoke API still works without changing model requests. At least two instances and serving-framework prefix caching are needed. Overloaded instances can spill requests elsewhere. AWS's reported latency improvement depends on its specific shared-prefix benchmark, not every prompt.

Watch next
Compare cache-hit rate and latency in a SageMaker routing benchmark using consistent request serialization.
Source evidence for SageMaker routes shared prompts to warm caches
Source-linked proof · original source

Video edition · Markdown edition

Source ledger