Model-selection guide · text generation
There is no single best model, and a ranking is only worth trusting if you can see the prompts behind it. This guide gives you the six properties that decide the choice, what to weigh for common writing tasks, and a short test to run on your own prompts.
Illustrative example
A made-up brief to show the shape. It is not a test result.
What this page is, and is not. ZeroTwo sells a multi-model workspace, so we have an interest in you comparing models. This page does not rank models, quote benchmark scores or publish test results, because we have not published a test set you could check. The task guidance below is editorial opinion, and the model table is read from ZeroTwo's own catalog. It was rewritten in October 2026.
What decides the choice
Model names change every few months; these properties do not. Decide which ones your task cares about, then check them on the candidates.
Start from the job, not the leaderboard. Write the task as one sentence plus the exact input you will give the model. A model that writes warm prose can be the wrong tool for bulk classification, and the reverse is true too.
CheckJudge candidates on your own input, never on a demo prompt.
A stated maximum is a ceiling. Three numbers matter: the provider's maximum, the limit the product exposes, and how much the model reliably recalls.
CheckHide facts mid-document and ask for them back.
Input and output caps are separate. A model that reads a whole report may still stop a long draft early.
CheckCompare the maximum output tokens, not only the context.
Some models always reason before answering, some let you toggle it, some do not. Reasoning usually costs latency and extra output tokens.
CheckTry with thinking off first, then on for the failures.
Providers price input and output tokens differently, and prices change. In a subscription, the plan sets the limit instead.
CheckRead the provider's current price page; do not trust a copy.
A model licence and a product's data policy are different documents. The licence says what you may do with the weights or outputs. The data policy says what the service keeps, for how long, and whether it trains on your inputs. Hosted models are governed by the provider's terms; open-weights models by their licence plus wherever you run them.
CheckRead both before you put client or personal data in.
Editorial guidance
This is our reading of what usually matters, not a benchmark. It names no winning model on purpose: the right candidate depends on your input, so the last column tells you how a task usually goes wrong.
| Task | Weigh most | Try first | Failure to watch |
|---|---|---|---|
| Long-form drafting | Output cap, instruction following, a consistent voice | Thinking off. Ask for an outline, approve it, then draft section by section. | Repetition and drift from the brief after the first few sections. |
| Editing and rewriting | Faithfulness to the source, tone control | A fast model with thinking off and a short style sample in the prompt. | Silent changes of meaning; over-polished text that loses your voice. |
| Summarising long documents | Context you have verified, recall of details | The largest context you have tested; ask for quotes with page references. | Skipped middle sections and details that are not in the source. |
| Structured extraction | Format compliance, repeatable output | Thinking off, an explicit schema, the smallest model that passes your set. | Valid JSON with wrong values. Validate against the source, not the schema. |
| Multi-step analysis and planning | Reasoning quality, showing intermediate work | Thinking on at a moderate level; ask for assumptions before conclusions. | A confident wrong intermediate step, and the cost of long reasoning. |
| Coding help | Tool use, repository context, test-fixing loops | Start from the dedicated guide, which covers this task on its own. Choosing a model for coding. | Code that runs but does not do what was asked: run the tests. |
| Short replies at volume | Latency, cost per call, consistency | The smallest model that passes your test set; escalate only the failures. | A quality cliff on rare or ambiguous inputs. |
| Creative writing and roleplay | Voice, instruction following, how it handles dark themes | Give three models the same opening and compare, blind if you can. | Samey phrasing, and refusals that differ by provider. |
From ZeroTwo's catalog
These are the text models ZeroTwo's bundled catalog marks as main models, with the figures it lists. They show how much the same property can differ between models. The catalog can lag the live list, so check the model catalog or AI models news for what is current. Context is the figure listed here, not a tested recall limit, and provider maximums can differ.
| Model | Provider | Listed context | Thinking | Released |
|---|---|---|---|---|
| GPT 6.1 Sol | OpenAI | 1.05M | Always thinks | September 2026 |
| GPT 4.1 | OpenAI | 1M | No thinking step | April 14, 2025 |
| GPT 5.2 | OpenAI | 400K | Toggle on or off | December 2025 |
| Claude Sonnet 4.6 | Anthropic | 200K | Toggle on or off | February 2026 |
| Claude Sonnet 4.5 | Anthropic | 200K | Toggle on or off | September 29, 2025 |
| Claude Haiku 4.5 | Anthropic | 200K | Toggle on or off | October 15, 2025 |
| Gemini 3.1 Pro | 1M | Toggle on or off | February 2026 | |
| Gemini 3 Flash | 1M | Toggle on or off | November 18, 2025 | |
| Grok 4.1 Fast | xAI | 2M | No thinking step | November 20, 2025 |
| Grok 4.1 Fast R | xAI | 2M | Always thinks | November 17, 2025 |
| DeepSeek V4.1 Flash | DeepSeek | 1M | Toggle on or off | September 10, 2026 |
| Kimi K3.1 Pro | Moonshot (Kimi) | 1M | Always thinks | October 24, 2026 |
| Kimi K3.1 Pro Highspeed | Moonshot (Kimi) | 1M | Always thinks | October 24, 2026 |
| Kimi K3 | Moonshot (Kimi) | 1M | Always thinks | July 15, 2026 |
| Kimi K2.5 | Moonshot (Kimi) | 256K | Toggle on or off | January 26, 2026 |
Source: ZeroTwo's own model catalog as bundled with this site. Prices are left out on purpose: they change, so read each provider's current price page.
Test it yourself
Public rankings average over tasks that are not yours. A small test on your own prompts, scored the same way each time, tells you more.
One sentence for the task, the exact input you will give it, and what a good output looks like.
Vary one property at a time: a larger and a smaller model, thinking on and off, or two providers.
Include a normal case, a long one and an awkward one. Save each prompt with the model ID and the date you ran it.
Hide which model wrote which output and score each against the rubric. A colleague who has not seen the labels is better still.
Providers update models and retire IDs. Re-run the set when a model you rely on changes.
| Criterion | How to score it (0 to 5) | Starting weight |
|---|---|---|
| Correctness against the source | Count claims that are unsupported or wrong. Zero tolerance for facts you will act on. | 30% |
| Instruction following | Did it respect length, audience, constraints and the requested structure? | 20% |
| Format compliance | Does the output parse and match the schema or template every time? | 15% |
| Voice and style | Does it read like something you would publish after light editing? | 15% |
| Latency | Time to a usable answer for your real prompt, measured the same way on each. | 10% |
| Cost per task | Provider price for the tokens used, or the plan allowance the task consumed. | 10% |
Running the comparison in ZeroTwo
ZeroTwo lists 60+ models from several providers in one workspace. Plans differ in how many of them you can use. Figures below come from the same source as the pricing page. Pro is $26.99 per month billed annually.
$0/mo
$14.99/mo
$29.99/mo
$120/mo
Questions
There is no single best model, and this page does not rank them. A model that suits long-form drafting can be slow or costly for bulk extraction, and the order changes with every release. Write your task as one sentence, pick three candidates, run the same five to ten prompts on each, and score the outputs against the rubric on this page. That test is more reliable than any list, including ours.
Open-weights models let you control where the model runs and how it is tuned, and you take on the hosting, scaling and licence reading. Hosted models trade that control for convenience. Read the specific licence before you fine-tune, redistribute or ship a model in a paid product. A chat workspace such as ZeroTwo gives you hosted access to models; it does not hand you the weights.
Estimate it from your real input. As a rough rule of thumb for English, one token is about three quarters of a word, but use the provider's own tokenizer for a real count. Then separate three numbers: the provider's stated maximum, the limit the product you use exposes, and how much the model reliably uses. Only the first two are published. Measure the third by hiding a few facts in the middle of a long input and asking for them back.
ZeroTwo is a multi-model workspace, so you can send the same prompt to different models and compare the outputs. This page does not claim how much of a conversation carries over when you switch models partway through a thread. To find out, start a long chat with an attached file and a standing instruction, switch models, and check whether earlier messages, the file and the instruction are still used.
ZeroTwo lists 60+ models from providers including OpenAI, Anthropic, Google, xAI and DeepSeek; the current list is on the models page. Model selection depends on your plan: Free ($0) starts with 100 free credits and 20 bonus credits every day you log in, with limited model selection, Plus ($14.99 per month) offers unlimited fast models, Pro ($29.99 per month, or $26.99 per month billed annually) a better selection, and Ultra ($120 per month) all models.
Yes. Every large language model can state wrong facts fluently, most often on niche topics, recent events and exact figures. Give the model the source text to work from, ask it to quote the passage it relied on, and check any fact you will act on against the original. A better model reduces the problem but does not remove it.
Privacy depends on the product and the model provider, so check both. ZeroTwo's practices are described in its privacy policy, and each provider has its own data terms for the models it hosts. Do not paste confidential material into any tool until you have read the terms that apply to your plan.
Keep going
Start on the Free plan, no card required, and read the current plan terms before you decide.