Several Models in One App: Routing, Benchmarks and Native Tool Costs in T3 Code

T3 Code's FAQ says it is not a harness and does not work without one. Each provider needs its own install and login: the Codex CLI and codex login, Claude Code and claude auth login, the Cursor CLI on a paid Cursor plan, the Grok Build CLI and grok login, OpenCode 1.14.19 or newer, Antigravity with its own Google sign-in, and fx on the machine that runs the server. Pi, Muse Code and agents from the ACP Registry can be added too.

Its bundled model manifest, updated Oct 28, 2026, lists Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 for Claude, and GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna for Codex, which otherwise takes its model list from its app server. Through Antigravity the model list comes from the user's Google account, including Claude and GPT models when the plan has them. API keys and OpenRouter work as an extra provider instance with its own keys and endpoint. The FAQ says T3 Code does not support local models.

One prompt can go to several models, each in its own thread and git worktree, and T3 Code transfers conversation context when a thread switches providers. Through its orchestrator MCP server an agent can create a sub-agent on any supported provider instance and wait for the result; the child gets only the task prompt, not the parent's history. An agent can also hand a task to a linked machine with delegate_task.

Scheduled tasks run on a chosen environment with its project, model and workspace settings, and a webhook schedule can start one when GitHub opens a pull request or a CI job fails. Agents T3 Code did not start, such as Claude Code or Codex in a terminal, can work with a T3 Code environment through its MCP server. Browser tabs belong to the environment, and a Device panel shows an iOS Simulator or Android Emulator that agents control.

On Vals.ai's Terminal-Bench 4.0 run of Oct 24, 2026, pass@1 averaged over 3 runs in the mini-swe-agent harness, Claude Opus 5.5 scored 65.2, Claude Sonnet 5.5 64.1 and Kimi K3.1 Pro 63.6, with GPT-6 Astra at 59.6 and Gemini 4 Argon at 57.6. Sonnet 5.5 lists at US$2 input and US$10 output per million tokens against Opus 5.5's US$4 and US$20.

Other dimensions rank models differently. On Artificial Analysis' AutomationBench-AA, a tool-use test, Gemini 4 Argon scored 77.5 against Kimi K3.1 Pro's 73.3. On AA-LCR, a long-context reading test, Kimi K3.1 Pro scored 86.3 and Claude Fable 5.1 85.3, but Kimi K3.1 Pro measured 41 output tokens per second against Claude Haiku 5.5's 241. The matrix on the T3 Code page sets these side by side with every value linked to its run.

Practitioners split work by role more than by vendor. The pattern with the widest support is a frontier model that plans and a cheaper one that implements. One Hacker News commenter plans with Opus through the API, does the rest with DeepSeek or GLM Flash and runs review loops with Sol 6.1, putting a very complex task at about US$10 at most (Oct 17, 2026). Another, after exhausting a Claude Max 20x plan in about two days, moved simple work to Haiku 5.5 (Oct 19). A third reports cutting Claude costs at work from about US$200 to about US$50 a month with Opus planning, Haiku coding and Sonnet reviewing (Oct 31).

Keep more than one model family in rotation: the leaders change often, and each family catches different problems in review. Theo Browne, who builds T3 Code, codes with Opus 5.5 and reviews with GPT-6.1 Sol; in his Oct 26, 2026 video: "It grabs onto the problem and doesn't let go, and I really like using it for reviews for this reason." Unattended runs need that second reader. Armin Ronacher, co-founder of Earendil, writing on his blog on Sep 7, 2026 after a 35-hour unattended GPT-6 Astra run: "It has shown that it will commit slop, and it requires me to review it more as a result."

Views on Gemini 4 Argon for code split sharply. The developer who builds T3 Code would like to auto-close pull requests to its repository filed using Gemini 4 Argon. Theo Browne, in his Oct 14, 2026 video testing Gemini 4: "I don't trust any Gemini models nor Muse Wonder. ... I would like to find some way to auto close all pull requests that are filed using models like Gemini 4 Argon or the normal Muse Wonder when they are filed to T3 code." One commenter reports Argon forgetting significant parts of detailed instructions on mid-sized real tasks (Oct 14), while another's eval of 11 small bug fixes had it solve 10, at a median US$0.52 a run against US$0.18 for Sonnet 5.5 (Oct 13).

A reviewer from another model family catches what the author's model misses. Simon Willison, the independent developer behind Datasette, writing on his blog on Oct 13, 2026: "I've been trying adversarial review in my own projects recently, where you have a model try and find flaws (I often play Claude and GPT off against each other)." A second-model review before a PR opens adds time up front and tends to return PRs with fewer issues that merge faster, and it puts an idle subscription to work.

Budget the loop: three rounds, the first full, later rounds only on changed code and open findings, more only for security or data-loss risk, and a stop when a round returns only nits. Reviewers from other model families can push over-engineering and scope creep without that cap. One commenter saw Codex with Astra heavily refactor a one-off script it was only asked to review (Sep 28).

Give each agent its own worktree; parallel work is far easier to manage that way, and T3 Code gives each model selected for one prompt its own thread and worktree. T3 Code's model picker chooses a harness as well as a model, so a skill set up in Claude Code is not there when a thread moves to Codex. Set skills up in each harness a team uses.

Name the sub-agents in the prompt, since the parent model chooses them: for example, ask a GPT-6.1 Sol thread to start two Opus 5.5 delegates to explore performance wins, then a new thread on Grok 4.7 to implement the findings. A sub-agent can run on any supported provider, in its own harness and subscription. Expect plans to drain faster, because far more threads run at once.

In T3 Code, Claude runs through Anthropic's Agent SDK, so test Claude-heavy work in T3 Code before moving a team onto it.

Some model APIs include server-side tools that reach data sources which otherwise need their own paid API. xAI's API prices its X Search tool at US$5 per 1,000 posts, billed per post fetched, including parent and quoted posts, plus the model's tokens. X's own API sells post reads at US$0.005 each from prepaid credits, the same per-post rate with no model attached.

Gemini's API takes a public YouTube URL as video input. Google lists the feature at no charge and says pricing and rate limits are likely to change, with a free-tier cap of 8 hours of YouTube video a day, and, in static mode, about 100 tokens per second of video at default (low) resolution. The YouTube Data API gives a default 10,000 quota units a day and 100 search.list calls, but captions.download requires permission to edit the video, so it cannot fetch another channel's captions.

A coding agent can reach X data by calling Grok, whose xAI API includes X search. Theo Browne, who builds T3 Code, in his Oct 14, 2026 video testing Gemini 4: "if I want to get good data out of Twitter, I can't do it through the API, it's way too expensive. So, I'll just tell Codex or Claude to call Grok, and by calling Grok, it gets access to the XAI APIs."

On the prices GILD stored on Nov 1, 2026, xAI's X Search and the X API's pay-per-use post reads both cost US$0.005 per post, while xAI also bills token usage for the model used in the request.

For web search, Gemini's Grounding with Google Search, on the Gemini API's paid tier, gives 5,000 free requests a month across Gemini 3 and newer models, then US$14 per 1,000, with a default cap of 1,500 requests a day. Anthropic's web search tool costs US$10 per 1,000 searches and OpenAI's US$10 per 1,000 calls, each plus tokens; xAI's costs US$5 per 1,000 calls. Brave's Search API sells results at US$5 per 1,000 requests with US$5 of free credits a month, and Google's Custom Search JSON API is closed to new customers and ends on Jan 1, 2027.

Read the routes against your volume. For public YouTube videos, Gemini's URL input is the only route here that reads the video itself, since the Data API cannot download another channel's captions. For X posts the two routes cost the same per post, so the model route pays off when you want a model to read the posts anyway. For search, Gemini's 5,000 free grounded requests a month cover light use, but past them it is the dearest per search at US$14 per 1,000 with a 1,500-a-day cap, while xAI's tool and Brave's API both list US$5 per 1,000.

Three kinds of access are easy to confuse. A model API's own tool, such as X Search or a YouTube URL, is billed by the model vendor when your code calls that API. An integration inside T3 Code is whatever the agent in a thread exposes, its own tools and the MCP servers configured for it, billed by whatever sits behind them. An external service, such as the X API or the YouTube Data API, is a separate account with its own price and quota.

T3 Code's usage view shows token use and estimated API-equivalent cost per harness, and its FAQ says those estimates are not the subscription bill.

OpenAI documents Sign in with ChatGPT for open-source and locally hosted apps and documents Codex app-server for that route; paid or remotely hosted apps must complete an interest form. That route does not support image generation, file search, Code Interpreter, native computer use or hosted MCP connectors.

Anthropic's legal page says products built on the Agent SDK should use API key authentication, and that third-party developers may not route requests through Free, Pro, Max or Ultra plan credentials on behalf of their users; it does not prevent an end user from signing in to the unmodified Claude Code binary. Anthropic's support page, updated June 15, 2026, says Agent SDK and third-party app usage still draw from subscription limits. T3 Code's FAQ says it is within Claude Code's terms and that its Agent SDK route drains 5-hour limits faster, which it calls a bug on Anthropic's side.

This is GILD's first-party account and counts toward no practitioner theme. GILD runs its research agents in T3 Code, with Claude Opus 5.5 and GPT-6 Astra each checking work the other produced. In review and QA the two were complementary: Astra's check rejected 417 of 1,287 page additions Claude Opus 5.5 builders wrote, and Claude's checks of three of Astra's fact-backing jobs accepted 75 of 114 rows; GILD discarded all four Astra backing jobs in that batch for false dates.

Astra underperformed at extracting claims. On the same 67 sources it coded 93 claims to Opus's 123, and its support levels matched the final codes on 53 of 82 claims against Opus's 74 of 105. GILD discarded Astra jobs that falsified dates or paraphrased sources, and since November 1 Astra only fetches and crawls pages while Claude writes every row.

GILD read T3 Code's README, docs, model manifest and release history, the agents' own pages, and the xAI, X, Google, OpenAI, Anthropic and Brave pricing and docs pages on Nov 1, 2026, and stored each fetch. Benchmark figures are the independent runs GILD stores for each model, with their evaluator and date. Prices are list prices on the date read. GILD's own figures come from its research records.

• T3 Code+20 PositiveAI Coding Agent by T3 ToolsAdd to My Stack

Official documentation

19

Public source

30

1. 1T3 Code docs: FAQOfficial documentationgithub.com
2. 2T3 Code READMEOfficial documentationgithub.com
3. 3T3 Code model manifest (updated Oct 28, 2026)Official documentationgithub.com
4. 4T3 Code internals: Model manifestOfficial documentationgithub.com
5. 5T3 Code docs: Thread sidebarOfficial documentationgithub.com
6. 6T3 Code docs: Portable handoffsOfficial documentationgithub.com
7. 7T3 Code docs: Orchestrator MCP serverOfficial documentationgithub.com
8. 8T3 Code docs: Remote accessOfficial documentationgithub.com
9. 9T3 Code docs: Project settingsOfficial documentationgithub.com
10. 10T3 Code docs: Outside agents over MCPOfficial documentationgithub.com
11. 11T3 Code docs: DevicesOfficial documentationgithub.com
12. 12T3 Code docs: UsageOfficial documentationgithub.com
13. 13T3 Code docs: CodexOfficial documentationgithub.com
14. 14T3 Code docs: ClaudeOfficial documentationgithub.com
15. 15T3 Code docs: CursorOfficial documentationgithub.com
16. 16T3 Code docs: OpenCodeOfficial documentationgithub.com
17. 17T3 Code docs: AntigravityOfficial documentationgithub.com
18. 18T3 Code docs: fxOfficial documentationgithub.com
19. 19T3 Code docs: ACP agentsOfficial documentationgithub.com
20. 20egeozcan on Hacker News (Oct 17, 2026)Public sourcenews.ycombinator.com
21. 21A_D_E_P_T on Hacker News (Oct 19, 2026)Public sourcenews.ycombinator.com
22. 22surgical_fire on Hacker News (Oct 31, 2026)Public sourcenews.ycombinator.com
23. 23necovek on Hacker News (Sep 28, 2026)Public sourcenews.ycombinator.com
24. 24Revanche1367 on Hacker News (Oct 14, 2026)Public sourcenews.ycombinator.com
25. 25ariwilson on Hacker News (Oct 13, 2026)Public sourcenews.ycombinator.com
26. 26AI Coding Daily (YouTube, Oct 19, 2026)Public sourceyoutube.com
27. 27Simon Willison: Redesigning source control at Anthropic (Oct 13, 2026)Public sourcesimonwillison.net
28. 28Armin Ronacher: Astra for Coding: Why Are We Doing This Again? (Sep 7, 2026)Public sourcelucumr.pocoo.org
29. 29Theo Browne: These 3 skills changed my life (YouTube, Oct 26, 2026)Public sourceyoutube.com
30. 30Theo Browne: So I've been testing Gemini 4 (YouTube, Oct 14, 2026)Public sourceyoutube.com
31. 31Theo Browne: So I've been testing Gemini 4, at 14:25 (YouTube, Oct 14, 2026)Public sourceyoutube.com
32. 32Vals.ai: Terminal-Bench 4.0 leaderboard (Oct 24, 2026)Public sourcevals.ai
33. 33Artificial Analysis: Gemini 4 Argon (read Oct 27, 2026)Public sourceartificialanalysis.ai
34. 34Artificial Analysis: Kimi K3.1 Pro (read Oct 27, 2026)Public sourceartificialanalysis.ai
35. 35Artificial Analysis: Claude Haiku 5.5 (read Oct 28, 2026)Public sourceartificialanalysis.ai
36. 36xAI API pricing: tools (read Nov 1, 2026)Public sourcedocs.x.ai
37. 37X API pay-per-use pricing (read Nov 1, 2026)Public sourcedocs.x.com
38. 38Gemini API: Video understanding (read Nov 1, 2026)Public sourceai.google.dev
39. 39Gemini API pricing (read Nov 1, 2026)Public sourceai.google.dev
40. 40YouTube Data API: Quota costs (read Nov 1, 2026)Public sourcedevelopers.google.com
41. 41YouTube Data API: captions.download (read Nov 1, 2026)Public sourcedevelopers.google.com
42. 42Claude API: Web search tool (read Nov 1, 2026)Public sourceplatform.claude.com
43. 43OpenAI API pricing: tools (read Nov 1, 2026)Public sourcedevelopers.openai.com
44. 44Brave Search API pricing (read Nov 1, 2026)Public sourcebrave.com
45. 45Google Custom Search JSON API overview (read Nov 1, 2026)Public sourcedevelopers.google.com
46. 46OpenAI: Sign in with ChatGPT for open-source apps (read Nov 1, 2026)Public sourcedevelopers.openai.com
47. 47OpenAI: Sign in with ChatGPT limitations (read Nov 1, 2026)Public sourcedevelopers.openai.com
48. 48Anthropic: Claude Code legal and compliance (read Nov 1, 2026)Public sourcecode.claude.com
49. 49Anthropic support: Use the Claude Agent SDK with your Claude plan (read Nov 1, 2026)Public sourcesupport.claude.com