The short answer
There is no single best AI model. The right pick depends on the job, and it changes every month. This page shows Spezia's current pick for each kind of work, such as coding, research and content, with the reason, the cost and what to avoid. The picks are rewritten every week from release notes, published results, pricing and hands-on findings, and nobody pays to be listed.
- The best pick depends on the job, so each pick names a model and the tool it runs in.
- Spezia rewrites the picks every week from release notes, published results and pricing.
- Each pick shows what it costs and what to avoid, because the wrong setup wastes money or time.
- Nobody pays to be listed. Vendors have no say in the picks.
- The table of recent changes shows new models, features, prices and limits from official release notes.
What is the best AI model right now?
It depends on the job. A model that leads on coding may not lead on research, and the tool wrapped around a model changes how well it works. So each pick below names both: the model and the harness it runs in. The table is this week's guide, read live.
Rewritten every week from release notes, published results, pricing and hands-on findings. Nobody pays to be listed.
Claude Opus 5.5 is now the first genuinely publishable AI writer and the strongest software model, while OpenAI's cheaper small models (GPT-6.1 Terra) take over browser automation and high-volume agent work.
Software engineering and coding agents
Claude Opus 5.5 in Claude Code
- Why
- Expert hands-on finding that Claude is much better than alternatives for software engineering, and Claude Code's release notes show active investment in agent task management, spend tracking and latency tuning. Cost: $4/$20 per 1M input/output tokens
- Watch out for
- Don't run unattended agents on live production without controls: frontier agents still execute 40.64% of dangerous OS-level operations on LITMUS, and LinuxArena exists precisely to test sabotage risk.
Browser use and web automation
GPT-6.1 Terra in ChatGPT Work / Codex
- Why
- Expert finding that OpenAI models are more token-efficient than Claude for browser use, and Terra is explicitly designed for fast, efficient computer use at scale with a 75% cost cut versus Haiku 4.5. Cost: 75% less than Haiku 4.5
- Watch out for
- Don't default to Claude for browser automation right now, expert notes say OpenAI is more token-efficient for this job.
Content and media: writing, design, video
Claude Opus 5.5 in Claude apps (Pro/Max/Team)
- Why
- Expert finding that Opus 5.5 is the first model where AI-produced content is genuinely publishable; Max and Team plans now include monthly API credits for running content agents. Cost: $4/$20 per 1M input/output tokens
- Watch out for
- Don't rely on older small models for publishable writing, the expert note specifically credits Opus 5.5 as the first genuinely publishable tier.
Outreach and sales: email, LinkedIn, CRM
GPT-6.1 Terra in ChatGPT Work / Codex
- Why
- Expert note directs browser-run outreach to OpenAI for token efficiency, and Terra is the cheap high-volume computer-use model. Cost: 75% less than Haiku 4.5
- Watch out for
- Don't use Claude for browser-driven outreach, expert notes say OpenAI is more token-efficient there.
Data, spreadsheets and finance work
Claude Haiku 5.5 in Claude Platform
- Why
- Cheapest, fastest, most capable Claude small model for high-volume tasks. Cost: about 75% cheaper to run than the older Haiku
- Watch out for
- Don't lock in a model on price alone, Mistral Large 4's blended price doubled and then halved inside one week.
Read each row from left to right. The first column is the kind of work. The second is the pick. The third is why it wins, with the cost where it is known. The last is what to avoid, which matters as much as the pick. Setups that waste money or time are named as plainly as the ones that work.
How are the picks chosen?
From evidence, and never from payment. Spezia reads release notes, published benchmark results, pricing pages and hands-on reports every week, and rewrites the guide from them. Labs report their own benchmark numbers, so Spezia shows where each result came from. Nobody pays to be listed, and no vendor has a say in a pick.
Two habits keep the guide honest. The pick names a harness as well as a model, because the same model can behave differently in different tools. And the guide names a runner-up and what to avoid, so you can see what was close and what was not.
What changed in AI tools in the last 30 days?
Tools change monthly, and a change in price or limits can undo a good pick. This table lists the most important recent change from each tool Spezia watches, read from the vendor's own release notes.
Read from each vendor's own release notes. Fixes and documentation edits are left out, and each tool appears once, with its most important change.
New models, price changes, new limits and deprecations rank above new features. A deprecation is the one to read first, because it has a date after which something stops working.
How do you choose a model for your own work?
Five steps, in this order.
- Name the job. Write it in one sentence. "Draft customer emails" and "review pull requests" are different jobs with different best picks.
- Start from the pick and the runner-up. Both are in the table above. Ignore everything else until these have been tried.
- Test on five real examples. Use your own work, not a demo. Compare the output, the time and the cost.
- Write down the cost. Price per task matters more than price per token, because a cheaper model that needs three tries costs more.
- Check again monthly. A new release or a price change can move the pick. The table above changes when it does.
How can your coding agent use this?
Spezia connects to the coding agents people already use, so they can look up the current pick for a job instead of working from stale memory. Connect it to Claude Code, Codex, Cursor, Gemini CLI or VS Code in one step from your account. Your agent keeps running on your own model and account. Spezia supplies the current facts.
Open the tooling guide to read the full reasoning behind each pick, with sources, and ask Spezia for exact setup steps.
See this week's picksHow this was made
The picks are read live from Spezia's weekly tooling guide, which is rewritten every week from release notes, published results, pricing pages and hands-on findings. Nobody pays to be listed. The list of changes is read from each vendor's own release notes, with one entry per tool, and fixes and documentation edits are left out. The tables on this page refresh within hours of each update.
Sources
Common questions
Which AI model is best for coding?
It depends on the tool it runs in and the kind of coding, and it changes often. The coding row in the table at the top of this page is Spezia's current pick, with the reason, the cost and what to avoid. It is rewritten every week.
How often does the best AI model change?
Often. New models, price changes and new limits arrive monthly, and any of them can move a pick. Spezia rewrites its guide every week and lists the most important changes from the last 30 days on this page.
Is Spezia paid by any AI vendor?
No. Picks come from release notes, published results, pricing and hands-on use. Nobody pays to be listed, and a vendor cannot buy a higher place in a pick.
What is a harness?
A harness is the tool wrapped around a model, such as a coding assistant or a research app. The same model can perform differently in different harnesses, so each pick names both the model and the harness it runs in.
Where do the benchmark numbers come from?
From the labs' own announcements and from independent reports, each shown with the words it was read from. Labs report their own numbers, so Spezia shows where each came from. Spezia does not run the benchmarks itself.
Cite this
Spezia. "The best AI model and tool for each job, updated every week" Spezia, November 10, 2026. https://spezia.ai/blog/best-ai-model-for-each-job
You may quote and share this analysis with a link to the post. Plain-text version · RSS