| Part | What it holds | A public description |
|---|---|---|
| Task | Goal, starting state, success rule | Anthropic: a test with inputs and success criteria. Harbor’s task overview stores that as instruction.md, task.toml, environment/, and tests/. |
| Isolated trial | A fresh machine or workspace | Anthropic: start from a clean environment. Inspect’s sandbox page says the Docker provider runs each sample in a container, with sample_init at the start of the sample. |
| Agent runner | The same loop users get | Anthropic: the agent in the eval should function roughly the same as the production agent. Harbor’s README evaluates agents such as Claude Code, OpenHands, and Codex CLI. |
| Transcript | Ordered log of the attempt | NVIDIA: user message, each step, environment state when the attempt stops. Anthropic also says trace or trajectory. |
| Grader | Code, a model rubric, or a person | Anthropic’s three types: code-based, model-based, and human. Harbor’s verifier writes a number to /logs/verifier/reward.txt or reward.json. |
| Aggregation | More than one trial, then a report | Google: do not block a pull request on one noisy run; watch the aggregate pass rate. NVIDIA: report a range across 3–5 trials, not a point estimate. |
| User agent | What OpenAI’s crawler page says it is for | What a Disallow does, according to that page |
|---|---|---|
| OAI-SearchBot | Surface websites in ChatGPT search features. | The site is not shown in ChatGPT search answers. It can still appear as a navigational link. |
| GPTBot | Crawl content that may be used in training generative AI foundation models. | A Disallow indicates the content should not be used to train those models. |
| ChatGPT-User | Visit a page when a person in ChatGPT or a custom GPT asks it to. | Not used to decide Search inclusion. The page says robots.txt rules may not apply because the fetch is user-initiated. |
| Date | What is on the record |
|---|---|
| October 28, 2026 | NVIDIA schedules the third-quarter fiscal 2027 call for November 17, 2 p.m. Pacific time. |
| November 4, 2026 | Mast, Moolenaar, and Crawford ask Huang for a closed-door briefing on November 18, and for an answer by November 11. |
| November 5, 2026 | The House Foreign Affairs Committee publishes that request. |
| November 6, 2026 | The Verge, citing the Times, describes a reported effort to learn about American AI chips and to influence sales of Nvidia’s most advanced chips to China. |
| November 7, 2026 | TechCrunch reports the memo claim and a later Financial Times report of a romantic relationship, plus an earlier NVIDIA comment that the interactions were “not unusual.” |
| November 8, 2026 | The board opens the investigation. NVIDIA publishes Huang’s email. NVIDIA tells CNBC he will continue as CEO without interruption. |
| Sunday night, reported by Axios on November 9 | Axios quotes Trump saying he has not heard that Huang did anything wrong, and that he would certainly take him on an Asia trip. |
| November 11, 2026 | The House letter’s deadline to accept November 18 or propose another session day. |
| November 17 and 18, 2026 | Earnings call, then the briefing date the chairs requested. |
| Long-input task | IAPO, one episode | First order only, one episode | Strongest baseline | IAPO, matched budget |
|---|---|---|---|---|
| GSM8K, 790 test problems | 78.1% at 5.6k tokens | 77.4% at 9.7k | LLM-Debate, 79.5% at 28.6k | 81.8% at 25.1k, +2.3 |
| MATH, 220 test problems | 37.7% at 5.6k | 37.6% at 15.2k | DyLAN, 48.3% at 54.3k | 48.6% at 50.8k, +0.3 |
| DROP, 800 test problems | 69.2% at 11.5k | 67.9% at 20.1k | LLM-Debate, 70.6% at 57.1k | 74.2% at 46.0k, +3.6 |
| MBPP, 333 test problems | 81.1% at 5.1k | 63.8% at 15.7k | DyLAN, 71.2% at 121.9k | 80.9% at 33.1k, +9.7 |
| HumanEval, 131 test problems | 83.7% at 9.9k | 76.5% at 18.1k | MaAS, 83.8% at 26.5k | 84.3% at 20.4k, +0.5 |
| Chatbot | RAG app | Fixed workflow | AI agent | Multi-agent system | |
|---|---|---|---|---|---|
| Who chooses the next step | The person does, by sending another message. OpenAI’s guide says a question-answering chatbot can call a tool for one current document and still cannot plan a multi-step job unless every step is programmed. | The application does. Lewis et al., 2020, retrieve passages and then a generator writes the answer conditioned on those passages. | The workflow does. OpenAI’s guide says the steps and the “if this, then that” rules are written in advance. An LLM may handle one judgment, such as classifying a request, and then control returns to the script. Microsoft’s sequential pattern is the multi-agent version of a fixed line: the next agent is defined by the workflow. | The model does. The harness runs the tool the rules allow. Anthropic’s SDK and OpenAI’s SDK both describe that pass from model to runtime. | A pattern does. OpenAI’s handoff gives the specialist the branch. An agent used as a tool leaves the manager in charge of the final answer. Microsoft also documents a fixed sequence, a parallel fan-out, a shared thread, and a manager with a task ledger. |
| Whether it acts | It replies in text. Anthropic’s loop page says that without tools, Claude can only respond with text. | It reads retrieved text into the answer. In the 2020 paper the retriever is a Wikipedia index, and the generator writes. A product that keeps choosing new searches has left that one pass. | It performs the actions named in the rules. OpenAI’s examples are a locked account after failed logins, and an expense flow that stalls on a blank field. | It can read, write, search, call an API, or run a command, inside the permissions. The model requests the action. The harness carries it out. | Each role can act inside its own tools. A2A is for one independent agent delegating to another. MCP is for an agent calling its own tools. |
| What “done” looks like | A message in the thread. | A generated answer that used the retrieved passages. Lewis et al. also treat the retrieved text as something a person can inspect. | The predetermined outcome, such as a routed ticket or a locked account. OpenAI’s guide says every branch was written down, which is why the result is auditable and why an unlisted condition can stall the run. | The goal is met, the model stops calling tools, a turn or budget limit hits, a guardrail fails the run, or a person has to approve before the run resumes. | The owning role returns the answer, or the pattern’s stop rule fires: a checker approves, a manager’s ledger is complete, or an iteration limit ends the thread. |
| Part | What it does |
|---|---|
| Goal | The task the run is trying to finish. Anthropic’s defined term starts the agent loop when the agent receives a task. OpenAI’s business guide says an agent starts from a goal and can pause to ask for clarification. |
| Model | The component that reads the current context and proposes the next step or a final answer. OpenAI’s guide says the model interprets instructions, plans, and decides what to do next. Anthropic’s SDK sends that proposal back as text, tool calls, or both. |
| Instructions and policies | Standing text and checks. OpenAI’s business guide separates guardrails, which bound the action, from instructions, which guide the decision. The Agents SDK uses guardrails for automatic checks and a separate approval pause before a side effect. Anthropic’s permission mode, allow list, and disallow list are the same kind of boundary on the loop page. |
| Tools | The actions the model may request. The harness validates and runs them. MCP is one open connector. A single programmed lookup, in OpenAI’s chatbot example, is a tool call that still leaves the plan in the application’s hands. |
| State and memory | What the next turn can see. Anthropic resends conversation history and can keep a session. OpenAI’s running-agents guide offers four ways to carry state: replayed history, a session your application stores, a conversation id, or the previous response id. The business guide says memory of past steps lets an agent refine a later pass, and it describes that memory as still developing. |
| Harness and control loop | The software that feeds the goal, offers tools, stores observations, enforces the rules, and decides when to call the model again. On this site that software is the harness, and the cycle it runs is the agent loop. Anthropic’s SDK overview says the SDK gives you the same tools, agent loop, and context management that power Claude Code. The overview also names Managed Agents as a hosted harness that runs the loop in an Anthropic-managed cloud sandbox or a self-hosted sandbox. |
| Stop or escalation | The run ends on a final answer with no further tool calls, on a turn or budget limit, or on a failed guardrail. It pauses when a person must approve. OpenAI says to resume that pause from the saved state. A person who approves before the action counts is human in the loop. A person who watches the record and can step in is human on the loop. |
| Kind of data | Tasks | What OpenAI says it includes |
|---|---|---|
| Gels and blots | 31 | Agarose DNA gels, SDS-PAGE gels, Western blots |
| Protocols, notes, and specs | 27 | Protocols and SOPs, lab notebooks and run records, photographed handwritten notes, catalogues, instrument specs, published methods |
| Tables and instrument exports | 20 | CSV and Excel exports, lane layouts and inventories, raw flow cytometry (FCS) files |
| Bench, apparatus, and reagent photos | 18 | Benches, gel rigs, reagent and culture tubes, a PCR plate, tube racks, pH strips, a concentrator, a freezer, a liquid-handler deck |
| Microscopy | 14 | Raw brightfield and fluorescence images, multi-channel sets, z-stacks, time-lapse movies |
| Instrument screens and plots | 13 | Photos of flow cytometer, microscope, NanoDrop, plate reader, and balance screens; exported standard curves, chromatograms, and plots |
| Sequencing files | 12 | Raw long-read FASTQs, FASTA sequences, GenBank maps |
| Protocol videos | 7 | Streaking, inoculation, serial dilution, SDS-PAGE tank and incubator setups, weighing on a balance |
| Agar plates and colonies | 7 | Transformation plates, blue/white screens, antibiotic-disc plates |
| Reference sheets | 21 | Ladder charts, plasmid maps, pH charts. OpenAI marks these as supporting material. |
| Model | Mean reward | 95% CI | Output tokens per task |
|---|---|---|---|
| GPT-6 Astra | 66.7% | 60.3–73.1% | 4,704 |
| GPT-6.1 Sol | 62.0% | 55.6–68.4% | 5,238 |
| GPT-5.6 Sol | 43.8% | 37.6–50.0% | 9,099 |
| GPT-6 Luna | 38.3% | 32.3–44.5% | 13,120 |
| GPT-5.5 | 35.4% | 29.7–41.0% | 11,135 |
| GPT-5.4 | 33.2% | 28.3–38.4% | 13,717 |
| GPT-5.2 | 31.2% | 26.2–36.3% | 27,620 |
| Chat | Fixed automation | Workplace agent | |
|---|---|---|---|
| What you hand it | A question, a rewrite, or a short draft. | A trigger and a rule, such as “lock the account after three failed logins.” | A goal and the sources it may use. ChatGPT’s Work guide says the useful tasks have a clear outcome and something you will review. |
| Who chooses the next step | The conversation stays on the reply. OpenAI’s guide says a tool call for a lookup is still not a plan. | The workflow does. A model may be allowed one step, such as classifying a request, and then control returns to the script. | The model proposes the step. The harness runs the tool it is allowed to run. |
| When a fact is missing | It answers from what it has, or it asks you. | OpenAI’s guide says a fixed expense flow can stall on a blank field. | The guide says an agent can try another source or ask a person, then continue. |
| What “done” looks like | A message in the thread. | The predetermined outcome, such as a routed ticket. | A file, a scheduled run, or an action in a connected app, held for review when the product requires it. |
| Where this showed up on November 7, 2026 | Claude chat, and ChatGPT’s Chat mode. Claude’s Cowork page says a simple question still gets a quick answer. | The workflow examples in OpenAI’s business guide. Deloitte’s 2025 essay calls some “agent” projects this pattern under a new name. | Claude Cowork and ChatGPT Work. Shared workspace agents are a research preview on top of ChatGPT for selected plans. |
| Theme in the November 4 press release | What the release says it covers |
|---|---|
| Harness engineering at scale | Prompts, tools, memory, guardrails, and feedback loops become a product with owners, versions, and dependencies, plus evaluation, security, access control, and maintenance. |
| Putting more agents to work, without losing human judgment | Parallel sessions and longer tasks, including unattended overnight builds, can turn review and delivery into bottlenecks. The release warns against “human in the loop theater,” where approval is a ritual. The radar page heads this theme “Putting more agents to work.” |
| Beyond frontier models | A proprietary frontier model is no longer the obvious default. Open-weight models have narrowed the gap for many workloads and can run on an organization’s own infrastructure. Choose by task, cost, privacy, and operations. |
| The economics of running AI at scale | Cost includes routing, batching, storage, networking, and scheduling, not only the model price. The radar page heads the matching section “Constructing the AI factory.” |
| Part of the harness, in the November 4 post | The job it does |
|---|---|
| Instructions and skills | What the agent loads before it acts. |
| Tools | What the agent is allowed to call. |
| Permissions and sandboxes | The box the agent runs inside. |
| Context policy | What the agent is allowed to see. |
| State | What the agent carries between steps. |
| Orchestration | How the steps are sequenced. On this site, orchestration means choosing the next step or specialist. |
| Evaluation | Whether the run was effective. |
| Telemetry | What happened when it was not. |
| Blip | Ring in Volume 35 | What the radar text uses it for |
|---|---|---|
| Agent sandboxing | Adopt, techniques | A sensible default: isolated environments, restricted privileges, limited access to other systems, shaped by threat modeling. |
| Threat modeling | Adopt, techniques | A regular human-led exercise. The text says prompt injection and autonomous action can slip past rule-based scans. |
| Agent plugin marketplace | Trial, techniques | Distribution for skills, instructions, MCP servers, and subagent definitions, with a way to publish and update them. |
| Feedback flywheel | Trial, techniques | A team habit: ask at the end of a session what should change in a shared artifact, then decide in a retrospective. People still accept the edit. |
| Self-improving harness | Trial, techniques | Agents read traces and eval feedback, then propose or make changes to prompts, tools, skills, memory, and workflows. Humans set the bounds. |
| Risk-tiered human oversight | Trial, techniques | Review depth follows what a change could break, so attention goes where a person can still judge the diff. |
| Harness quality evaluation | Assess, techniques | Scheduled checks that a skill, plugin, or guardrail actually works. |
| Enterprise harness | Assess, techniques | The shared layer coding agents consume, kept thin and federated. |
| Human-in-the-loop theater | Caution, techniques | An approval step the person cannot realistically evaluate. The text calls it an accountability sink. |
| AI-accelerated shadow IT | Caution, techniques | Software built outside the supported path, including with coding agents. The suggested answer is a useful enterprise harness and firm security limits. |
| Adopt blip | Quadrant |
|---|---|
| Agent sandboxing | Techniques |
| Mutation testing | Techniques |
| Path-to-production mapping | Techniques |
| Post-quantum cryptography | Techniques |
| Progressive context disclosure | Techniques |
| Threat modeling | Techniques |
| VictoriaMetrics | Platforms |
| mise | Tools |
| OpenAI Codex | Tools |
| OpenCode | Tools |
| Open Data Contract Standard (ODCS) | Languages and frameworks |
| At $20 a month | Claude Pro | ChatGPT Plus |
|---|---|---|
| Where the price is printed | Claude pricing. Annual billing is $17 a month, $200 up front. Max and Ultra are monthly only. | Codex pricing docs. The ChatGPT pricing FAQ lists Go and Plus as monthly plans. |
| Coding agent | Claude Code is included. The pricing table says No on Free and Yes on Pro, Max, and Ultra. | Codex on the web, in the CLI, in the IDE extension, and on iOS. The docs call this enough for a few focused coding sessions a week. |
| What else is named for coding | The same subscription covers Claude on the web, desktop, and mobile. Terminal, IDE, Slack, and the web are the steering surfaces on the product page. | Cloud integrations such as automatic code review and Slack. GPT-6.1 Sol and GPT-6 Luna. Credits if you want to go past the included allowance. |
| How the limit works | One pool for Claude and Claude Code. Pro has a five-hour session limit and a weekly limit. Anthropic says there is no fixed message count. | Codex shares usage with ChatGPT Work. The docs print local-message estimates for a five-hour window, and say those estimates are not fixed limits. Weekly limits may also apply. |
| Plan | Price on November 6, 2026 | Coding agent on that plan |
|---|---|---|
| Claude Free | $0 | Claude Code is not included. The feature table says No. |
| Claude Pro | $20 a month, or $17 a month with annual billing ($200 up front) | Included. Shared with the rest of Claude. Five-hour and weekly limits. |
| Claude Max | $100 a month, billed monthly. The help center says this is the web price and that mobile prices can differ. | Included. 5 times Pro usage every five hours and every week. $100 a month in API credits on the Claude Platform. Fable, for up to half the weekly limit. |
| Claude Ultra | $200 a month, billed monthly. Web price; mobile prices can differ. The help center says the old Max 20x plan was renamed Ultra, with no price change. | Included. 10 times Pro usage every week. Verified accounts have no five-hour limit. $200 a month in API credits. Fable, for up to half the weekly limit. |
| ChatGPT Free | $0 | Codex for quick tasks. The help center includes Free, and says Codex Cloud is not included. The docs name GPT-6 Luna at Standard speed in the desktop app, subject to rollout. |
| ChatGPT Go | $8 a month | Codex for lightweight tasks. Codex Cloud is not included. Same Luna note as Free, subject to rollout. |
| ChatGPT Plus | $20 a month | Codex on web, CLI, IDE, and iOS, plus the cloud integrations named above. Eligible for Codex Cloud, subject to rollout and settings. |
| ChatGPT Pro | $100, $200, or $500 a month. The card says “from $100.” | Everything in Plus, on the docs page. Ultrafast mode is on the $500 plan. The docs say Pro has no five-hour limit. The Codex pricing page says GPT-6 Astra sits inside the allowance on the $100 and $200 Pro plans. |
| API key | API token rates, not a monthly sticker | Codex in the CLI, SDK, or IDE extension. The docs say this path has no cloud features such as GitHub code review and Slack. |
| Seat | Price | What the coding agent includes |
|---|---|---|
| Claude Team, standard | $20 per member per month billed annually, or $25 billed monthly. Minimum two members, up to 150. US prices, tax excluded. The product page prints the $20 annual figure and “up to $500 per month in pooled API credits.” | Claude Code on every seat. The help center says a standard seat includes 1.25 times the Pro plan’s per-session allowance, plus a weekly limit. Limits are per member. Monthly API credits on Team plans cannot be used in Claude Code or the Claude apps. |
| Claude Team, premium | $100 per member per month billed annually, or $125 billed monthly | Claude Code, with 6.25 times the Pro plan’s per-session allowance and a weekly limit, on the help center page. The pricing page says premium seats give 5 times more usage than standard seats. Both sentences are on Anthropic’s pages. This guide does not merge them into one formula. |
| Claude Enterprise | The pricing FAQ says $20 per seat per month plus usage at API rates, billed annually. The product card lists security, analytics, role-based access, spend limits, and data controls, and it does not repeat the dollar figure. | Included on new and self-serve Enterprise seats. The help center says usage-based Enterprise, including self-serve, has no per-seat usage limit and is billed at API rates. Older Enterprise plans can still split seats. |
| ChatGPT Business | $20 per user per month for two or more users, billed annually. $25 per user per month billed monthly. | ChatGPT and Codex on desktop and mobile. Larger virtual machines for cloud chats. Credits to extend usage. SAML SSO and MFA. No training on business data by default, on the docs page. |
| ChatGPT Enterprise and Edu | Contact sales. The docs card does not print a dollar price. | Everything in Business, plus priority processing, SCIM, retention and residency controls, audit logs, and the Compliance API, as that card lists them. On the docs page re-read November 9, 2026, flexible-pricing workspaces scale with credits and have no fixed rate limit. Enterprise and Edu plans without flexible pricing have the same per-seat usage limits as Plus for most features. |
| If you want | Lower list price on November 6, 2026 | What you give up at that price |
|---|---|---|
| To try a coding agent | Codex on ChatGPT Free, $0 | Claude Code is not on Free. Codex Cloud is not on Free. The docs describe quick tasks, not a full workday. |
| A small paid plan | Codex on Go, $8 a month | Still no Codex Cloud. Claude has no published plan between Free and Pro on the pricing page checked here. |
| The usual individual plan | Claude Pro at $17 a month if you pay annually. Month to month, both are $20. | Annual Pro is a year up front. Plus stays a monthly plan in the ChatGPT pricing FAQ. |
| A heavier individual plan | The $100 and $200 stickers match: Max or Ultra, and Pro at $100 or $200. | The multiples differ. Max is 5 times Pro on a five-hour and a weekly window. The Codex docs say Pro has no five-hour limit. Ultrafast is the $500 Codex tier. |
| A team, billed annually | Both entry seats are $20 per person, for two or more people. | Claude premium seats jump to $100 annually. A Codex FAQ line also names a $100 Business plan for fuller GPT-6 Astra access. Enterprise on both sides needs the sales or FAQ page, not this sticker. |
| Behavioral evaluation | End-to-end benchmark | |
|---|---|---|
| Question | Did this expected action happen? | Did the task succeed under the suite’s scoring rule? |
| What you read | A tool call, a file change, a question, or another step in the trace. | A score on a frozen task set. The harness is part of the bundle being scored. |
| What a drop tells you | Which behavior moved. Google’s examples are a missed question, a skipped validator, or a missing link. | That the bundle did worse. Google names Terminal-Bench and DeepSWE as suites where a small score change is hard to explain. |
| How strict it should be | Strict when one action is the right milestone. For a task with several correct paths, Google says to avoid a rigid tool sequence and use an outcome check, including an LLM-as-a-judge. | The suite’s own grader. A search over whole harnesses, such as HarnessFlow, lives at this end of the scale. |
| What it is for | Iteration and regression checks while you edit the harness. | A destination measurement. Google says behavioral evals do not replace it. |
| Evaluation set | Baseline passes / runs | Updated agent | Tool calls, before to after |
|---|---|---|---|
| Development, 2 tasks × 3 runs | 3/6 | 6/6 | 10 to 6 |
| Held out, 2 tasks × 3 runs | 4/6 | 6/6 | 28 to 15 |
| Product or path | Who runs the environment | Availability when the page was read | Notable limits |
|---|---|---|---|
| Claude Cowork and Claude Code | Your Mac or Windows computer. The desktop app has to be awake and running. | Research preview for Claude Pro, Max, and Ultra. Turn it on in the desktop app settings. Dispatch can hand the task off from the phone. | Anthropic says it asks for explicit permission first, and requests permission before a new app. Some apps are off-limits by default. Computer use is still early next to coding and text. Keep sensitive data out. |
| ChatGPT desktop app | Your Mac or Windows desktop, in the apps you allow. On Windows it uses the active desktop. | In supported regions, with ChatGPT Work and Codex. Open Settings, then Computer use, and turn on Any App. If that row says Install, follow the prompts. | Permission before each new app. It cannot automate a terminal app or ChatGPT itself. On Windows it takes the foreground. Stay present for payments and credentials. |
Claude API, computer_toolset_20260801 |
Your application, in an environment you control. Anthropic recommends a virtual machine or a container. | Messages API, re-read November 7, 2026. The quick start needs no beta header. The availability table lists computer_toolset_20260801 as available for Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, Claude Mythos 5, Claude Mythos Preview, Claude Opus 5.5, Claude Opus 5, Claude Sonnet 5.5, Claude Sonnet 5, Claude Haiku 5.5, and Claude Opus 4.8 on the Claude API and Google Cloud. A second available row, without Mythos Preview, covers AWS, Amazon Bedrock, and Microsoft Foundry. The beta row lists Claude Haiku 4.5, Claude Opus 4.7, Claude Opus 4.6, Claude Sonnet 4.6, Claude Opus 4.5, and Claude Sonnet 4.5. Not available in Claude Managed Agents. |
Seventeen member tools. You perform every call. Confirm financial transactions and other consequential actions. Treat the screen as untrusted. |
OpenAI API, ComputerToolset and BrowserToolset |
Your driver, on a browser or desktop you run, or a partner sandbox such as Daytona or E2B Desktop. | API guide re-read November 7, 2026. Samples use gpt-6-astra and gpt-6.1-sol. The page says you need a model that supports the computer tool. |
Isolate the environment. Confirm purchases, destructive changes, and typing secrets. Bound the steps, the time, or the cost. |
| OpenAI Agents API, hosted browser | OpenAI hosts the browser. The computer-use page points to a separate Agents API guide for this path. | That separate guide. The computer-use page does not list plan names for it. | The safety section on the computer-use page still says to confirm hard-to-reverse actions and to bound the run. |
| Gemini API Computer Use | Your client. Google recommends a sandboxed virtual machine, a container, or a separate browser profile. | Read November 7, 2026, at the AI Studio address. The older ai.google.dev address redirected there. A request on November 6 had returned HTTP 500. The changelog, re-read November 7, still dates general availability to October 30, 2026, on the Interactions API and the GenerateContent API. The model list names Gemini 4 Argon, Gemini 3.8 Flash as the recommended model, Gemini 3.5 Flash-Lite, and Gemini 3 Flash Preview. |
Browser, mobile, and desktop for Gemini 4 and Gemini 3.x. The client executes the action. Payments can be blocked or sent for confirmation. Google says the feature can contain errors. |
| Source, checked November 5, 2026 | Input | Output | Cache |
|---|---|---|---|
| TokenHub catalog, per million tokens | 6 yuan | 18 yuan | 0.3 yuan, cache hit |
| TokenHub bill through December 31, 2026, 23:59:59 Beijing time | 50 percent of catalog | 50 percent of catalog | 50 percent of catalog |
| Tencent’s English announcement, approximate catalog | US$0.90 | US$2.70 | US$0.045, cached input |
| OpenRouter, Tencent Cloud list, then the 50 percent rate shown on the page | $0.834, then $0.417 | $2.501, then $1.251 | $0.042, then $0.021, cache read |
| Capability | Hy4 on TokenHub | Hy4 Preview on TokenHub | Where Tencent describes it |
|---|---|---|---|
| Context window | 1M, with 960k input and 64k output | The same limits | TokenHub model list, November 4, 2026 |
| Image understanding | Listed | Absent from the row | TokenHub, and the announcement’s “first time” sentence for visual understanding |
| Function calling | Listed | Listed | TokenHub. See tool use. |
| Retained thinking | Listed, as 深度思考（保留式思考） | Listed | TokenHub’s label. The thinking from a step stays available. |
| Computer use | Named in the November 4 announcement | The announcement assigns this training claim to Hy4 | Command line and graphical interface, with Blender and CAD named as training software. TokenHub’s rows use the function-calling label instead. |
| Model, in Tencent’s financial evaluation | Score |
|---|---|
| Hy4 | 3.04 |
| DeepSeek-V4.1-Flash-1017 | 2.97 |
| Kimi K3.1 Pro | 2.68 |
| GLM-5.3 | 2.68 |
| Assistant | What the page says happens at purchase | Who sells | Where, and the source date |
|---|---|---|---|
| Google AI Mode and the Gemini app | Universal Cart holds items from more than one store. Checkout can use Google Pay in the cart, or the items can move to the merchant's cart. In Canada, direct checkout is rolling out on eligible retailers. | The retailer | U.S. cart rollout on October 15, 2026, in Search and the Gemini app. Canada direct checkout announced November 4, 2026. |
| ChatGPT | People compare products in the chat. Purchases finish on the merchant's website or app. OpenAI says it is moving away from a standalone Instant Checkout. | The merchant | Merchant page re-read November 7, 2026: shopping is live for ChatGPT users in the U.S. The course change is the March 24, 2026 post. |
| Alexa for Shopping | It can add items to a cart for you to check out, buy when a price you set is reached, or place Buy for Me on an eligible product from another store. | Amazon, for Amazon-store orders. The brand, for Buy for Me delivery and support. | Amazon's guide updated October 29, 2026: U.S. customers, in the Shopping app and on Amazon.com. |
| Microsoft Copilot | Copilot Checkout completes the purchase inside Copilot, on the merchant's existing payments, fraud, tax, and fulfillment systems. | The merchant | Product page re-read November 7, 2026: Copilot.com and the Copilot mobile app. English-language merchants selling to U.S. buyers, in U.S. dollars. |
| Perplexity | Instant Buy sends the order and the payment to a compatible merchant. PayPal's launch announcement powered that checkout. | The merchant | Help article retrieved November 4, 2026. It did not load on November 7, so this cell was not re-checked. Users in the United States, except Enterprise Pro and Enterprise Max. Launch announcement: November 25, 2025. |
| Product | What Google said on November 3, 2026 |
|---|---|
| Gemini | Off for K-12 student accounts until an administrator turns it on. Named in the blog and in the X clarification. |
| Gemini Notebook | Same default as Gemini. Named in both the blog and the X post. |
| Gemini in Classroom | On the blog’s list of four products that stay off until an admin turns them on. Not named in the X post’s list of three. |
| Google Search | Off by default. The X post says a school can turn it on without turning the other named services on. AI Mode and AI Overviews will offer step-by-step guided learning. |
| Part | What it is for | A public description |
|---|---|---|
| Tools | Let the model ask for a search, a file edit, a command, or an API call. The harness runs the call and returns the result. | Anthropic’s Agent SDK overview lists built-in tools for reading, writing, editing, commands, and web search. See tool use. |
| Sandbox | Give those commands a bounded machine so a bad step stays inside it. | OpenAI’s sandbox guide separates the harness, which it calls the control plane, from the sandbox, which it calls the execution plane. |
| Files and memory | Keep notes, plans, and progress that outlast one context window. | Anthropic’s November 26, 2025 note uses a progress file, a feature list, and git. Pydantic documents a memory notebook. |
| Orchestration | Choose the next step, tool, or specialist, and who owns the reply. | OpenAI’s orchestration guide. Glossary: orchestration. |
| Guardrails | Allow, block, replace, or pause an action. | OpenAI’s guardrails page and Pydantic’s guardrail docs. Glossary: guardrail. |
| Checks before “done” | Require evidence, such as a test or a click-through, before a feature counts as finished. | Anthropic’s long-running notes, linked below, use a feature list and browser checks. |
| Traces | Record the calls, the cost, and the failures so a person can reconstruct the run. | Microsoft’s Agent Framework overview includes observability on its Harness Agent. OpenAI’s SDK overview points tracing at its observability guide. |
| Product | What the vendor’s docs say | Where the loop runs |
|---|---|---|
| Claude Agent SDK | Anthropic’s overview says the SDK gives you the same tools, agent loop, and context management that power Claude Code, including permissions, sessions, hooks, subagents, and MCP. | A Python or TypeScript process you operate. The SDK runs the Claude Code binary. |
| Anthropic Managed Agents | The same overview calls Managed Agents a hosted agent harness. Sessions sit in an Anthropic-managed cloud sandbox or a self-hosted sandbox. | Anthropic’s service, or a sandbox you host. This is a different product from the Agent SDK. |
| OpenAI Agents SDK | The sandbox guide says the harness owns the agent loop, model calls, tool routing, handoffs, approvals, tracing, recovery, and run state. The sandbox is the machine that reads files and runs commands. | Your application, in the TypeScript or Python SDK. The SDK overview says the Agents API is the managed harness in OpenAI’s service. |
| Codex | The Codex platform post says one open-source harness powers the app, the CLI, and the IDE extension. It gathers context, uses tools, applies sandbox and approval rules, and carries work forward. Model access stays separate from that open-source layer. | The Codex app, CLI, or IDE, or your own app through the Codex SDK, codex exec, or app-server. |
| Microsoft Agent Framework | The overview describes a Harness Agent for long, multi-step tasks: planning and todo tracking, context compaction, file access and memory, don’t-ask-again tool approval, and observability. Graph and function workflows are the path when the steps are fixed. Microsoft calls the framework the successor to Semantic Kernel and AutoGen, and says that if you can write a function for the task, do that instead of using an agent. | Your app. The overview shows Python and .NET, and says the Go SDK is in public preview and requires Go 1.26 or later. |
| Google ADK 2.0 | The ADK 2.0 page says the release moves from a hierarchical executor to a workflow runtime. Agents, tools, and functions are nodes in a graph. It also describes dynamic workflows and collaborative workflows with a coordinator and subagents. | Your process. The page says ADK 2.0 is available for Python, TypeScript, and Go. It dates general availability to May 19, 2026 for Python, June 30, 2026 for Go, and August 21, 2026 for TypeScript. |
| Pydantic AI Harness | The harness docs say every Pydantic AI agent already has a light harness: a typed loop, a model, tools, and structured output. The harness package adds a workspace, a plan, memory, sub-agents, context management, and durable execution for long jobs. Guardrails can allow, block, replace, retry, or defer a tool call for approval. | Your process. The quick start says LocalWorkspace runs on your machine and is a sandbox only if you swap in one. The page checked November 4, 2026 names Modal, E2B, Render, Sprites, Daytona, and boxd. Its first example model string is still anthropic:claude-opus-5-5. The same section also shows a coder on gpt-6.1-sol. |