Free
- Unlimited text chats with GPT-6 (subject to abuse guardrails)
- Limited messages with uploads
- Limited deep research
- Limited Codex access
Compare models side by side
Choose two to four models and an effort level for each. Every number is a published test result with its source. Open any row to see who ran the test.
Comparing 3 models. Change any model or its effort level.
OpenAI
Anthropic
Google
Real work tasksclosest to everyday professional work
Agents' Last Exam v1.1
58.6%Developer reported
No published score
39.5%Developer reported
GDPval-AA v2.1 (Artificial Analysis)
1608.9Independent test
1866.3Independent test
No published score
Vals Legal Index v1
68.7%Independent test
72.3%Independent test
76.7%Independent test
Vals Finance Index v1
55.8%Independent test
65.6%Independent test
64.9%Independent test
PRBench Legal (Professional Reasoning Benchmark, Scale Labs)
47.9%Effort level not stated
47.6%Effort level not stated
49.4%Effort level not stated
Remote Labor Index (Scale Labs)
No published score
21.3%Effort level not stated
17.1%Effort level not stated
Knowledge and accuracyfacts, hard questions and making things up
GPQA Diamond (Graduate-Level Google-Proof Q&A, Diamond subset)
95.4%Independent test
90.6%Independent test
96.4%Independent test
Humanity's Last Exam
52.9%Independent test
61.4%Independent test
No published score
Humanity's Last Exam, with tools
No published score
67.7%Developer reported
No published score
SimpleQA Verified
73.9%Independent test
72.2%Independent test
76%Independent test
Reasoning and mathspuzzles, science and maths
ARC-AGI-2 (Abstraction and Reasoning Corpus, version 2)
94.2%Independent test
91.7%Independent test
91.3%Independent test
FrontierMath Tiers 1 to 3
93.7%Independent test
91.2%Independent test
86.0%Independent test
Agents and computer usedoing multi-step tasks with tools
ARC-AGI-3 (Interactive Reasoning Benchmark)
52.7%Independent test
No published score
46.3%Only published at medium
OSWorld 2.1 (offline set)
78%Developer reported
No published score
No published score
AutomationBench (Zapier)
36.1%Developer reported
40%Developer reported
51.3%Developer reported
Codingsoftware changes
SWE-Bench Pro (V2 public set)
No published score
No published score
92.2%Independent test
DeepSWE v1.1
71.9%Developer reported
No published score
77.9%Developer reported
Terminal-Bench 4.0
56.1%Independent test
59.6%Independent test
57.1%Developer reported
Images and documentsreading charts, images and diagrams
MMMU-Pro (Massive Multi-discipline Multimodal Understanding and Reasoning, Pro version)
86.0%Independent test
87.7%Independent test
No published score
What people preferblind votes between two answers
LMArena Text Leaderboard (now hosted at arena.ai)
1481Independent test
1509Only published at high
1520Independent test
Scores come from different testers using different set-ups, so small gaps mean little. Developer-reported scores have not been checked by anyone else. Open a row to see who ran each test and how.
Does more effort help?
Higher effort lets a model work longer before it answers. It costs more and takes longer. These charts use ARC-AGI-2, a reasoning puzzle test whose tester publishes the score and the cost per task at each level.
Reported by ARC Prize Foundation.
Reported by ARC Prize Foundation.
Reported by ARC Prize Foundation.
Plans and prices
Pick an app to see its plans. The lowest monthly price for one person is shown on each button.
ChatGPT plans
A general assistant with projects, memory, a built-in browser, deep research and an agent mode. Codex coding access comes with every plan.
A$13 a monthOne price, not per person.
- More messages with tools
- More uploads
- Longer memory
- Limited deep research
A$30 a monthOne price, not per person.
- Expanded messages and uploads
- Expanded deep research
- Expanded Codex usage
- Codex local messages per 5 hours (estimates): GPT-6 Astra 5-45, GPT-6.1 Sol 15-160, GPT-6 Luna 350-3,000
A$155 a monthOne price, not per person.
- Your choice of 3 usage tiers
- Pro plans currently have no five-hour limit (Codex)
- Up to A$770 in monthly API credits (on the top tier)
- weekly GPT-6 Pro message limits in Chat (amount not stated)
A$70 a monthFor 2 people (minimum 2).
- For teams of 2-200 employees
- Unlimited subject to abuse guardrails
- 'Starting in December, plan usage will have a shared limit across all of ChatGPT and Codex'
- A$28 API credits per paid Standard seat, pooled up to A$770 per workspace
A$350 a monthFor 2 people (minimum 2).
- More usage than Standard, with no 5-hour limit
- Dots, always-on agents for Premium seats
- A$155 API credits per paid Premium seat, pooled up to A$770
What ChatGPT connects to
- Apps and connectors (connected apps). Use apps on all plans; create and share them from Go.
- Business connectors (connected apps). Business plans.
- File uploads (files and documents). Limited on Free; included from Go.
- ChatGPT Work (agent) (agents and tasks). Limited on Free and Go (desktop app); desktop, web and mobile on Plus and Pro.
- Scheduled tasks (scheduled work). Go and above.
- Projects (projects). All plans.
- Memory (memory). Limited on Free; expanded on Plus and Pro.
- Built-in browser (browser use). All plans.
- Deep research (research). Limited on Free and Go; included on Plus and Pro.
- Dots (agents and tasks). Always-on agents, Pro and Business Premium seats.
Good to know
- Scheduled tasks from the Go plan up.
- Built-in browser and projects on every plan.
- Five GPT-6 family models, most with six effort levels in the API.
- Plus usage limits for Codex are published as ranges, not fixed numbers.
- The Go plan may include ads.
- The AUD price for the higher Pro tiers is not shown.
Prices in US dollars are converted at the Reserve Bank of Australia rate of 0.6972 US dollars to the Australian dollar on 4 Nov 2026. Your card's exchange rate, fees and any GST change what you pay.
Building your own tools? API prices are separate
Developer access is billed per million tokens (a token is a short chunk of text), on a separate account from any subscription. Higher effort levels produce more tokens, so the same task costs more at max effort.
| Model | Input, per million tokens | Output, per million tokens | Notes |
|---|---|---|---|
| GPT-6 Astra | US$10 (about A$14.34) | US$50 (about A$71.72) | Standard processing, context under 272K tokens. Batch processing is half price. Source for GPT-6 Astra (opens in a new tab) |
| GPT-6.1 Sol | US$2 (about A$2.87) | US$10 (about A$14.34) | Source for GPT-6.1 Sol (opens in a new tab) |
| GPT-6 Luna | US$0.10 (about A$0.14) | US$0.50 (about A$0.72) | Source for GPT-6 Luna (opens in a new tab) |
| Claude Fable 5.1 | US$10 (about A$14.34) | US$50 (about A$71.72) | Source for Claude Fable 5.1 (opens in a new tab) |
| Claude Opus 5.5 | US$4 (about A$5.74) | US$20 (about A$28.69) | Batch processing saves 50%. Read from the US-dollar view; the Australian view shows the same figures marked as including GST, with no currency code. Source for Claude Opus 5.5 (opens in a new tab) |
| Claude Sonnet 5.5 | US$2 (about A$2.87) | US$10 (about A$14.34) | Source for Claude Sonnet 5.5 (opens in a new tab) |
| Claude Haiku 5.5 | US$0.10 (about A$0.14) | US$0.50 (about A$0.72) | Prompts up to 100K tokens. Longer prompts cost more. Source for Claude Haiku 5.5 (opens in a new tab) |
| Gemini 4 Argon | US$2 (about A$2.87) | US$10 (about A$14.34) | Rises to US$4 input and US$20 output from 1 December 2026. Output includes thinking tokens. Source for Gemini 4 Argon (opens in a new tab) |
| Gemini 3.8 Flash | US$0.75 (about A$1.08) | US$3.75 (about A$5.38) | Rises to US$1.50 input and US$7.50 output from 1 January 2027. Source for Gemini 3.8 Flash (opens in a new tab) |
| Grok 4.7 | US$2 (about A$2.87) | US$6 (about A$8.61) | API usage is billed separately from X Subscription plans. Source for Grok 4.7 (opens in a new tab) |
What a configured system involves: the Grokbot example
An example of what a configured AI system involves, as distinct from buying a plan. A fleet of Grok Bot agents works from written standards and routines held in a private repository. Each bot has a defined role, and a person approves anything that is published.
Built from
- Grok Bot platform access on a legacy Cursor Pro+ plan.
- 15 bots and about 31 scheduled routines.
- Written standards and routines in a private repository the bots read.
- One image and video generation provider, billed per credit.
- An email inbox for requests between systems; approvals stay with a person on a separate review site.
Running costs
- Platform plan: Cursor lists Pro+ at US$60 a month (legacy plan page, checked 4 Nov 2026).
- Generation caps: Recorded caps of US$1.00 per content edition and US$1.50 per recap.
- Spend tracking: Every paid call is written to a spend ledger and totalled in a weekly report.
- Actual monthly spend to date.
- Which model each bot runs: app defaults are accepted and not recorded.
- How much of the weekly platform allowance is used.
- Any AUD price for the platform plan.
Shown to illustrate the extra parts a configured system needs (standards, routines, spend caps, approvals and a ledger), not as a plan to buy. Its running cost depends on configuration and use, so it is not comparable with a subscription price.
What the scores mean for your work
What a benchmark is
A benchmark is a fixed test set. Developers and independent testers run a model on it and publish the share of tasks it gets right. Each one checks one narrow skill. None of them is built around Australian law, regulators or your organisation's procedures.
What effort means
Most current models can think for longer before answering. Developers call this effort, thinking or reasoning level. More effort is slower and costs more. It helps most on multi-step problems and least on quick rewrites or simple facts.
For workers compensation: 3 partly relevant, 10 general signal only, 7 not relevant. The list starts with the most relevant. Treat a high score as a sign of the underlying skill, never as proof a tool suits your work.
How this comparison works
This edition was checked on 4 Nov 2026. It holds 647 published scores for 33 models from 12 developers across 20 benchmarks. Prices were checked on 4 Nov 2026.
Where the numbers come from
- 350 scores come from independent testers such as ARC Prize, Epoch AI, Artificial Analysis, Vals AI, Scale and LMArena. The other 297 are reported by the developer and have not been reproduced by anyone else.
- Every score was checked a second time against its source page before it was published. Scores that could not be confirmed were left out.
- Each score keeps the effort level the source used. Where a source did not say, it is marked as not stated.
- Results that measure different things under one name, such as a newer version of a test, are kept apart or left out, so each row compares like with like.
What it deliberately does not do
- No overall winner or single score. Each model is stronger at some tasks than others.
- No guessed numbers. A blank means nobody has published that result.
- No converted prices where the developer already shows Australian dollars. Other prices use the Reserve Bank of Australia rate and are marked as estimates.
Six tasks to test an app on your own work
Benchmarks are run by others, under their own conditions. These six everyday tasks are a fairer test of fit. They are published so you can run them yourself on de-identified material. This edition has not run them yet, so no app is scored on them.
Cited researchAnswer a current question from public sources, with every claim linked to a source that actually says it.Not yet tested
- Test material
- Question: what changed in an Australian regulator's published guidance on a named topic in the last 12 months? The tester picks a topic with a known, dated answer before the run.
- How to mark it
- Every factual claim has a link, and the linked page supports the claim.
- Dates and version numbers match the source.
- Uncertainty is stated rather than filled in.
- No invented sources, titles or quotes.
- Write down for every run
- Product, plan and model; Effort or thinking setting; Search or research mode on or off; Date and time; Full output and every link checked.
Document comparisonFind every material difference between two versions of a policy document.Not yet tested
- Test material
- Two versions of a fictional 12-page leave policy, with 15 planted changes: 10 material (eligibility, timeframes, approvals) and 5 cosmetic.
- How to mark it
- Material changes found, out of 10.
- Cosmetic changes correctly labelled as cosmetic.
- No changes reported that are not in the documents.
- Each change cites the clause it came from.
- Write down for every run
- Product, plan and model; Effort setting; How the files were supplied; Date; Full output.
Spreadsheet analysisSummarise a dataset correctly and show the working.Not yet tested
- Test material
- A synthetic 2,000-row incident register (dates, sites, incident types, lost days) with three planted data-quality problems: duplicates, a mis-keyed date and a blank site field.
- How to mark it
- Headline totals match the answer key.
- All three data-quality problems flagged.
- Method shown (formula, code or steps) so a person can re-run it.
- No claims beyond what the data supports.
- Write down for every run
- Product, plan and model; Effort setting; Code execution or analysis tool used; Date; Output and any generated file.
Professional draftingDraft a plain-English letter from structured facts without adding facts or making the decision.Not yet tested
- Test material
- Facts for a fictional return-to-work plan update addressed to [CLAIMANT_NAME], claim [CLAIM_NUMBER], with a review date and the name of [TREATING_PRACTITIONER]. The decision itself is supplied; the model only drafts.
- How to mark it
- Every supplied fact present and unchanged; no new facts.
- Placeholders kept exactly as supplied.
- Plain English at the requested reading level.
- Decision language left to the human decision-maker.
- Write down for every run
- Product, plan and model; Effort setting; Any style instructions used; Date; Full output.
Small coding taskMake a small, testable code change and prove it works.Not yet tested
- Test material
- A 200-line script that calculates business-day deadlines, with one known bug around public holidays and an existing test file.
- How to mark it
- Bug fixed and the existing tests still pass.
- A new test that would have caught the bug.
- No unrelated changes.
- A correct explanation of the cause.
- Write down for every run
- Product, plan and model; Effort setting; Agent or chat mode; Date; Diff and test output.
Multi-step workflowComplete a chained task across files and steps, and stop for approval where a person must decide.Not yet tested
- Test material
- Read three synthetic meeting notes, build an action register, draft a summary email and stop before sending. A planted step asks for approval.
- How to mark it
- All actions captured with owner and due date.
- Summary email matches the register.
- Stops at the approval point and does not send.
- Steps and tool calls visible for review.
- Write down for every run
- Product, plan and model; Effort setting; Connectors or agent features used; Date; Full transcript.
General information and education only. Not legal, compliance, financial, procurement or professional advice. Test results are published by others and change often, as do prices and features. Check the linked source and your organisation's policies before choosing a tool, and use de-identified material in any test. Related: the AI Readiness Assessment and the resource finder.