Models, harnesses, and form factors. Compare the tradeoffs before choosing your setup.
Research checked September 4, 2026. Published capabilities support the analysis; real-world fit still needs a test.
Progress is saved in this browser.
Models, harnesses, and form factors. Compare the tradeoffs before choosing your setup.
Research checked September 4, 2026. Published capabilities support the analysis; real-world fit still needs a test.
Progress is saved in this browser.
Start with the distinction
Like choosing an engine, an operating setup, and a driver's seat.
What can it reason about or generate?
Examples: GPT-6 Astra, Claude Fable 5.1, Gemini 3.8 Flash.How does the work run?
Context, tool execution, state, permissions, and stopping rules.Where do you interact and review?
Chat, IDE, file workspace, browser, or an embedded service.One harness can use multiple models and appear in multiple interfaces. A comparison of apps is not a clean comparison of models.
Review the AI FundamentalsEvidence before preference
Primary documentation establishes published features and limits. Vendor benchmark results remain vendor reports.
The quadrants are Longhand's assessment. They are hypotheses to test, not measured scores or promised returns.
METR's 2025 randomized study found a slowdown for 16 experienced developers across 246 tasks. Its 2026 follow-up says the newer estimates are unreliable.
These findings do not establish the effect of today's models on your team.
Selected vendor model and harness families, plus five workplace form factors. Open-weight deployment, specialist models, pricing tables, and the full vendor market are outside this snapshot.
Model · Research checked Sep 4, 2026
OpenAI's new frontier model, with access rolling out in stages.
SWOT: Longhand analysisBroad reasoning, coding, and visual understanding make it a candidate for complex work across large collections of text and images.
The API model produces text and lacks native audio or video input; those workflows require additional models, preprocessing, or tools.
Template-heavy reporting and software maintenance offer measurable pilots for reducing the time needed to produce a reviewed deliverable.
Staged availability creates uncertainty for teams with near-term delivery commitments; verify actual access before making this model a workflow dependency.
Scope: Launch access is staged. Published performance comparisons are provider evaluations; tool-assisted scores also depend on the harness and evaluation settings.
Source basis: OpenAI documents a 1.05M-token context window, text and image input, and support for reasoning, coding, computer use, and document creation.
Model · Research checked Sep 4, 2026
Anthropic's generally available frontier model for long-running knowledge work.
SWOT: Longhand analysisLong-running reasoning, coding, and document vision make it a credible candidate for research and deliverables spanning many steps and source files.
Premium token rates and fallback on some biology and cybersecurity requests mean available capability and cost vary with the task.
Repeated, context-heavy workflows may benefit from cheaper cache reads when teams reuse substantial instructions, project history, and source material.
Enterprise retention requirements may exclude the default 30-day configuration; eligibility and future safeguard rollout dates can constrain adoption.
Scope: Benchmarks are provider-reported and can include fallbacks. Restricted Mythos results do not establish ordinary Fable capability; enterprise retention exceptions require eligibility.
Source basis: Anthropic describes stronger long-running reasoning and document vision. Cache reads cost $0.25 per million tokens; some biology and cybersecurity queries use fallback models.
Model · Research checked Sep 4, 2026
Google's generally available workhorse with adjustable reasoning and introductory pricing.
SWOT: Longhand analysisLarge context, adjustable reasoning effort, and low introductory token rates make it a practical candidate for repeated knowledge and software workflows.
Complex tasks can consume more tokens through extra reasoning and tool calls, particularly at higher effort settings; inexpensive tokens do not guarantee inexpensive results.
High-volume document and software workflows offer room to test whether lower token rates translate into lower cost per accepted result.
Announced token rates double in January 2027, changing workflow economics; introductory pilot savings should not become permanent budget assumptions.
Scope: Capability comparisons are Google's evaluations. Flash Cyber is a separate restricted variant; its results should not be attributed to ordinary Flash.
Source basis: Google documents a 1M-token context window, adjustable thinking, and introductory pricing through December 2026. Higher effort can consume more reasoning and tool tokens.
Harness · Research checked Sep 4, 2026
Coding workflows across terminal, editor, desktop, and cloud
SWOT: Longhand analysisLocal commands, scriptable workflows, and parallel cloud tasks connect implementation to existing engineering tools and review processes.
Cloud work needs an explicitly configured environment; local dependencies, tools, and credentials add setup work when moving tasks between environments.
Teams with maintenance backlogs and reliable CI can pilot parallel fixes against acceptance criteria they already use.
Malicious issues, dependency instructions, and websites can redirect connected agents toward unsafe changes or disclosure of code and secrets.
Scope: Cloud agent internet is off by default; setup scripts still have access. GitLab integration is Beta. Controls differ by client.
Source basis: The CLI reads files, edits code, and runs commands. Cloud tasks run in isolated environments with configurable dependencies and repository access.
Harness · Research checked Sep 4, 2026
An extensible Claude harness for repository and command-line work
SWOT: Longhand analysisRepository context and extensible tools support repeated investigation, coordinated edits, and verification across local, cloud, and remote workflows.
Bash isolation, file permissions, computer access, and unsandboxed exceptions require separate configuration; one sandbox setting does not cover every action.
Teams with established scripts and testing practices can package recurring engineering work into shared skills and integrations.
Untrusted repositories, dependencies, or integrations can introduce malicious instructions and code that exploit overly broad filesystem or network access.
Scope: Bash sandboxing supports macOS, Linux, and WSL2, not native Windows. File tools and computer use operate under different boundaries.
Source basis: Claude Code combines tools, context management, and execution environments. Skills, hooks, MCP, and subagents extend its gather, act, and verify loop.
Harness · Research checked Sep 4, 2026
Multi-step knowledge work across files and connected business tools
SWOT: Longhand analysisFiles, connected apps, and scheduled tasks support multi-step deliverables and let users review work across supported surfaces.
Local file and computer access depend on an online desktop app; cloud processing and surface differences complicate assumptions about where work happens.
Teams with recurring reports and scattered documents can pilot delegation where an experienced owner already knows what a correct deliverable looks like.
Malicious emails, documents, websites, and plugins can turn connected business access into unintended real-world actions.
Scope: Cloud and web/mobile are Beta on eligible paid plans. Local access needs desktop online; computer use has no screen sandbox.
Source basis: Cowork supports connectors, skills, plugins, files, and scheduled tasks. Cloud sessions continue across desktop, web, and mobile, subject to access limits.
Harness · Research checked Sep 4, 2026
A configurable Gemini harness for terminal and scripted workflows
SWOT: Longhand analysisScriptable terminal operation and configurable extensions connect model reasoning to existing local tools and repeatable project workflows.
Isolation requires careful profile selection: the documented default macOS Seatbelt profile restricts writes but permits broad reads and network access.
Organizations with reusable scripts and internal tools can introduce AI-assisted steps without replacing their whole development environment.
Third-party extensions and project hooks create supply-chain exposure; malicious content can trigger unexpected actions or disclose sensitive inputs.
Scope: Sandbox behavior varies by host and profile. Container isolation needs Docker or Podman; some orchestration features remain experimental.
Source basis: Gemini CLI uses local project context and supports shell tools, extensions, skills, hooks, MCP, and headless operation, with configurable sandbox methods.
Form factor · Research checked Sep 4, 2026
Discuss a question, choose sources, and review a written answer.
SWOT: Longhand analysisA familiar conversation supports clarification, source selection, and plan review, giving people several opportunities to steer the research before accepting its conclusions.
A polished report still requires source verification, while operational tasks need an additional handoff beyond the research workflow's documented read access.
Approved document stores and industry data services can connect research to the organization's evidence, making source access a useful investment alongside model choice.
Incomplete sources and changing access permissions can leave important evidence unavailable, so a confident synthesis may omit information the decision actually requires.
Scope: These sources illustrate one product's research workflow; they establish neither a current error rate nor features shared by every chatbot.
Source basis: ChatGPT deep research documents editable plans, source selection, and cited reports. Its launch documentation also identifies possible hallucinations and incorrect inferences.
Form factor · Research checked Sep 4, 2026
Work beside the code, commands, tests, and reviewable changes.
SWOT: Longhand analysisFiles, commands, diffs, and tests sit close together, giving the agent concrete feedback and giving a technical reviewer changes they can inspect.
Environment configuration and informed code review require expertise; incomplete tests and hidden repository assumptions limit what an apparently successful run can verify.
Existing test suites, continuous integration, and version control give teams infrastructure for bounded implementation pilots with reviewable outputs and explicit acceptance checks.
Malicious repository content and third-party tool responses can attempt to redirect the agent, making trusted inputs and constrained permissions relevant to development workflows.
Scope: Documented capabilities do not establish productivity gains; results depend on the repository, task, environment, tests, and the reviewer's technical judgment.
Source basis: Claude Code documents repository access, file editing, command execution, and IDE diffs. Its security documentation describes prompt injection risks and implemented safeguards.
Form factor · Research checked Sep 4, 2026
Delegate work on authorized files and inspect the resulting artifacts.
SWOT: Longhand analysisDirect access to working files lets an agent create usable artifacts, keeping the result in documents that colleagues can inspect, revise, and reuse.
File access broadens the consequences of mistakes; effective use requires clear task scope, understandable permissions, attention to changes, and a practical recovery path.
Organized document collections and approved team templates create opportunities for repeatable report preparation, provided people can check each artifact against its source material.
Untrusted documents and plugins can introduce malicious instructions, while sensitive files increase the potential exposure when the workspace grants more access than necessary.
Scope: Desktop access does not imply local-only processing. Permissions, retention, and recovery behavior differ by product and should be checked for the actual workspace.
Source basis: Cowork documentation describes reading and changing authorized files, task monitoring, backups, and prompt injection risks. Its current sessions connect cloud execution to desktop resources.
Form factor · Research checked Sep 4, 2026
Operate websites and applications through their visible controls and screens.
SWOT: Longhand analysisScreen interaction gives an agent a route into existing applications, potentially reaching workflows where a suitable purpose-built integration is unavailable to the team.
Visual targeting, tool choice, and multi-application steps can fail; latency and repeated attempts make some interactions cumbersome or unsuitable when precision is essential.
Legacy systems with limited APIs create opportunities for supervised information gathering and bounded administrative tasks, especially when speed is less important than access.
Changing interfaces and malicious page content can disrupt a workflow or redirect the agent, making assumptions about familiar screens unreliable over repeated runs.
Scope: These are documented limitations of one implementation, not measured failure rates for every browser agent. Application fit requires testing in its actual environment.
Source basis: Anthropic's computer use documentation lists latency, coordinate errors, tool-selection mistakes, and reduced reliability across multiple applications, alongside prompt injection risks and oversight guidance.
Form factor · Research checked Sep 4, 2026
Embed AI work inside an application or asynchronous business process.
SWOT: Longhand analysisSoftware can submit work asynchronously, track its status, and place outputs inside an existing process without requiring someone to maintain a live conversation.
The team owns integration, evaluation, and failure recovery; additional autonomous steps can add latency, cost, and compounding errors before anyone reviews the result.
Repeated intake and reporting workflows can justify shared integrations, explicit quality checks, and consistent escalation paths when acceptance criteria are clear enough to evaluate.
Provider changes, upstream outages, and shifting input patterns can invalidate deployment assumptions, requiring teams to revisit evaluations and maintain a way to stop work.
Scope: Background execution alone does not provide durable business-process orchestration. Scheduling, retries, duplicate prevention, quality checks, and operational ownership remain implementation decisions.
Source basis: OpenAI documents asynchronous response generation and status polling. Anthropic's engineering guidance describes feedback, stopping conditions, and the potential for autonomy to compound errors.
Turn the SWOT into a decision
Proposed pilot: compare your current process with two candidate combinations on the same small set of representative tasks.
Use the same inputs, permissions, task brief, and acceptance rubric. Include missing information and failure cases.
Record accepted-result quality, review and rework time, elapsed time, and total cost. Include tool calls and retries.
Name an owner and review date. Agree on minimum quality and acceptable effort before seeing results.
This is a suggested evaluation design. The sample size and thresholds must fit your work; they are not validated universal cutoffs.
Source register · 1 of 5
Checked September 4, 2026 · US Central. Undated documentation is a snapshot of the page at review.
OpenAI · Vendor source · September 3, 2026
OpenAI · Vendor source · Not stated
Anthropic · Vendor source · September 2026
Anthropic · Vendor source · September 1, 2026 announcement
Google · Vendor source · September 2, 2026
Google AI for Developers · Vendor source · Not stated
Links also appear beside each SWOT judgment. Recheck model access, product features, and terms before a purchasing or deployment decision.
Source register · 2 of 5
Checked September 4, 2026 · US Central. Undated documentation is a snapshot of the page at review.
OpenAI · Vendor source · Not stated
OpenAI · Vendor source · Not stated
OpenAI · Vendor source · Not stated
Anthropic · Vendor source · Not stated
Anthropic · Vendor source · Not stated
Anthropic · Vendor source · Not stated
Links also appear beside each SWOT judgment. Recheck model access, product features, and terms before a purchasing or deployment decision.
Source register · 3 of 5
Checked September 4, 2026 · US Central. Undated documentation is a snapshot of the page at review.
Anthropic · Vendor source · Not stated
Anthropic · Project documentation · Not stated
Google · Project documentation · 2026-04-17
Google · Project documentation · 2026-07-17
Google · Project documentation · 2026-04-10
OpenAI · Project documentation · Not stated
Links also appear beside each SWOT judgment. Recheck model access, product features, and terms before a purchasing or deployment decision.
Source register · 4 of 5
Checked September 4, 2026 · US Central. Undated documentation is a snapshot of the page at review.
OpenAI · Vendor source · 2025-02-02
Anthropic · Project documentation · Not stated
Anthropic · Project documentation · Not stated
Anthropic · Project documentation · Not stated
OpenAI · Project documentation · Not stated
Anthropic · Vendor source · 2024-12-19
Links also appear beside each SWOT judgment. Recheck model access, product features, and terms before a purchasing or deployment decision.
Source register · 5 of 5
Checked September 4, 2026 · US Central. Undated documentation is a snapshot of the page at review.
METR · Independent research · July 10, 2025
METR · Independent research · February 24, 2026
Links also appear beside each SWOT judgment. Recheck model access, product features, and terms before a purchasing or deployment decision.
Your next move
A capable model. A harness with the right tools and controls. A form factor that fits how your team works.
Which weakness would matter most in our workflow, and what evidence would change our choice?
Research snapshot: September 4, 2026 · US Central. Capabilities and access can change. This page is not updated automatically.