Which AI model to use? (Sept 2026)

Every AI lab says its newest model is the best one. Most of them are right about something. This page sets out, category by category, which of the September 2026 models actually leads on medical, coding, legal, cybersecurity, reasoning, images, video and more — with the published scores, the price, and the caveats the launch posts leave out.

The short answer

No single model wins everything. The right answer depends on the job.

The models compared

These are the current generally available flagships from each lab, plus the leading open-weight models. Claude Opus 5.5 is the first model in Anthropic's Claude 5.5 family; Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks. Claude Mythos 5.1 is deliberately absent: it is the same underlying model as Fable 5.1 with some safeguards lifted, and it is available only to vetted organisations in Anthropic's trusted-access programmes, not on any public plan or the standard API.

Category rankings

Reasoning and overall intelligence

The Artificial Analysis Intelligence Index v4.3 combines ten tests, including long-horizon knowledge work, business workflow automation, terminal tasks, document reasoning and factual accuracy, with 45% of the weighting on private test sets so vendors cannot train on them. The current ceiling is 58. Bars here are scaled to the leader.

If you have seen index scores of 57, 61 or 66 quoted for these models, those are older scales from earlier in September — Kimi K3, for example, launched at 57 on version 4.1, and Opus 5 at 61. Only v4.3 numbers are comparable with each other; the figures here are from the current v4.3.2 revision. Opus 5.5 was run at max effort with Anthropic's fallback models switched on, and trails on only three of the ten tests: CritPt, AA-LCR and GDP.pdf. Grok 4.7 was benchmarked on 21 September, the day it launched.

Science knowledge

GPQA Diamond is a set of graduate-level physics, chemistry and biology questions written so that experts outside the field cannot answer them with Google. The frontier is now bunched within six points of the ceiling.

Anthropic did not publish a GPQA Diamond score for Claude Opus 5.5 at launch.

Maths

FrontierMath Tier 4 is research-level mathematics. Note that these are OpenAI-reported figures and that OpenAI funded FrontierMath and had sight of many of its problems, so treat the gap as directional rather than settled.

Coding

The fairest single answer to "can this thing do engineering work" is now Artificial Analysis's Coding Agent Index. Each model runs inside its own maker's coding agent — Claude Code, Codex, Grok Build, Muse Code — on the same three test sets (DeepSWE, Terminal-Bench 4.0 and SWE-Atlas-QnA), scored independently.

Astra ties Fable 5.1 at roughly 60% of the cost per task, because it uses fewer tokens than any other agent on the index. Grok 4.7 gained nine points on Grok 4.6 to take fourth place, overtaking GPT-5.6 Sol. Claude Opus 5.5 had not yet been run through the Coding Agent Index at the time of writing. On Anthropic's own launch figures (vendor-reported), it scores 57.8% on CursorBench 4.0 against 51.8% for Fable 5.1 and 41.7% for GPT-5.6 Sol, and 54.4% on FrontierCode against 53.3% for GPT-6 Astra.

The older shell test is now saturated

Terminal-Bench 2.1, run by Artificial Analysis on the same harness for every model, was the index's shell test until 7 September. Ten models now score 84% or more, so it no longer separates the leaders, but it is still the widest independent comparison that includes the cheaper models.

The harder tests tell a different story

DeepSWE 1.1 puts a model inside a real repository and asks it to fix real issues. Terminal-Bench 4.0 is the much harder shell test that replaced 2.1 in the index. On these the order changes and the field spreads out.

Opus 5.5's vendor figure is at its xhigh effort setting, with a stated margin of error of about 2.6 points; on Artificial Analysis's independent run it is level with GPT-6 Astra and 11 points ahead of Opus 5. DeepSeek V4 Pro's 80.6% on the older SWE-bench Verified remains the best published open-weight result on that test. Grok 4.7 scored 73% on DeepSWE and 33% on Terminal-Bench 4.0 when Artificial Analysis ran it inside its own Grok Build agent on launch day.

What about PHP, C# and SQL?

Every headline coding score above is Python-biased: SWE-bench Verified is 500 tasks drawn from twelve open-source Python repositories, and DeepSWE and Terminal-Bench lean the same way. Nobody publishes a PHP or C# leaderboard. Two datasets get close, and both point the same way.

Opus 4.6 dropped only three points moving from Python-only Verified (80.8%) to Multilingual (77.8%), and on the private commercial set Claude degrades less than its competitors when moving from public to unseen repositories. Both figures are from an earlier Claude generation, because the current ones have not been re-run on these tests; the pattern is the point.

Desktop and computer use

OSWorld 2.0 gives a model a real desktop and 108 long workflows that take a human around 1.6 hours each. The headline numbers are partial-credit; the strict "finished the whole job" rate is far lower and shown alongside.

Strict end-to-end completion: Fable 5.1 41.7%, Opus 5 39.6%; no strict figure has been published for Opus 5.5. Anthropic's launch table, run on its current setup, puts Opus 5.5 at 81.8% against 80.7% for Fable 5.1 and 74.0% for Opus 5, so read its lead over Fable as about a point. Fable scores zero on any task where its safety classifiers intervene, which drags its average down on security-adjacent work. Astra completes tasks around 47% faster than GPT-5.6 Sol, so it often wins on wall-clock time despite the lower score.

Global knowledge

Humanity's Last Exam is 2,500 expert-written questions across every academic field. With tools enabled, Claude leads and this is the one academic test GPT-6 Astra loses.

Artificial Analysis's independent run, on its own setup, puts Opus 5.5 at 61.4% against a previous best of 59.1% from Fable 5.1, so the ordering holds. Gemini's figure is on the HLE-Verified variant, so treat it as approximate. Factual accuracy specifically is covered in the hallucination section below.

Biology and life sciences

This is the thinnest category. Only OpenAI publishes head-to-head life-science scores, Anthropic routes biology questions on Fable to Opus 5, and the strongest biology model on either side (Claude Mythos 5.1) is not on public sale: access is limited to vetted organisations in Anthropic's trusted-access programmes. Anthropic says the new Opus 5.5 matches or beats Mythos 5.1 across many areas of biology, which is why it ships with the same biology safeguards as Fable 5.1.

Legal (professional)

Legal benchmarks are the most fragmented, and each one is led by a different model. The consistent finding across all of them: models pass around 90% of the individual criteria on a legal task but complete very few whole tasks end to end.

Reasonable working rule: Claude for drafting and precision, GPT for research, and a human reads everything before it goes out. None of these models should be trusted to complete legal work unsupervised.

Cybersecurity

Two tests matter. ExploitBench measures whether a model can turn a known vulnerability into a working exploit; CyberGym measures whether it can find vulnerabilities in real codebases. A model can be excellent at one and mediocre at the other.

ExploitBench above. On CyberGym (vulnerability discovery) the order flips: DeepSeek V4.1 Flash 88.1%, Gemini 3.8 Flash cyber variant 86.2%, GLM-5.3 84.5%; Xiaomi reports 94.0% for MiMo-V2.6-Pro on its own setup, but only 47.9% on ExploitBench. Claude Fable 5.1 has no published cyber score because penetration-testing, exploit and binary-scanning requests are routed to Opus 4.8. Claude Opus 5.5, which Anthropic rates as comparable to Mythos 5.1 in cybersecurity, has no published score for the same reason: it is the first Opus to ship with Fable-class cyber safeguards. Gemini's cyber variant is only available inside Google's Fairwind programme, and Astra's public build declines advanced offensive work unless your organisation is enrolled in OpenAI's Daybreak scheme.

Tool use and agents

AutomationBench-AA is Artificial Analysis's version of Zapier's business-workflow test: multi-step tasks across real apps. The headline score gives partial credit; the strict figure counts only workflows finished without breaking a single rule.

Artificial Analysis says Claude Opus 5.5 is level with GPT-6 Astra on AutomationBench-AA, but its headline score was not in the launch summary, so it is not charted here. Strict, no-rule-broken completion: Astra 41.6%, Fable 5.1 32.1%, Opus 5 28.3%; Opus 5.5 scored 40.0% in Zapier's own strict run during early access, as published by Anthropic. On AA-Briefcase, the multi-week knowledge-work test, the order reverses: Opus 5.5 now leads at 1,822 Elo, 143 points ahead of Fable 5.1 and the first Anthropic model to beat GPT-5.6 Sol on presentation quality, with Fable 5.1 and Opus 5 behind it and Grok 4.7 joining the frontier just behind them, even though Grok 4.7 slipped 1.1 points against Grok 4.6 on AutomationBench. Grok 4.6 remains the value pick for automation: near the top at a fraction of Astra's price.

Medical and health

Three different questions get asked under "medical", and they have three different answers.

Professional use: answering to a physician's standard

HealthBench Professional is built from 5,000 conversations and physician-written rubrics from 262 doctors in 60 countries, scored on the length-adjusted version so that longer answers do not automatically win. It is OpenAI's benchmark, and all figures are OpenAI-run.

Fable 5.1 scoring below Fable 5 is a safety-routing artefact, not a capability regression: questions its classifiers flag are answered by a smaller model and marked down accordingly. No Gemini, Grok, open-weight or Claude Opus 5.5 score on HealthBench Professional has been published yet.

General medical information for the public

No consumer-grade health benchmark has been published for this generation of models, so the professional ranking above is the best available guide, read together with the hallucination table further down. The practical change this month is on the refusal side: Anthropic reports that Fable 5.1's biology safeguards fire 85% less often on benign elementary biology and medical questions than Fable 5's did, which means fewer "I can't help with that" answers to ordinary health queries.

Scan and imaging interpretation

The imaging benchmarks exist, but they live in academic papers rather than on vendor leaderboards: Radiology's Last Exam (RadLE) tests chatbots against board-certified radiologists on hard spot-diagnosis cases, ReXVQA covers chest X-ray question answering, CXR-LT is a multi-centre chest X-ray challenge with over 145,000 images, and newer sets such as NeuroQA cover 3D brain MRI. No lab publishes scores on any of them, so the only figures come from independent studies — which run a generation behind the current models.

Other studies agree on the ordering. On pneumothorax from chest radiographs, ChatGPT-4o was most accurate at 69.6%, then Claude 3.5 at 64.9% and Gemini 2.0 at 57.4% — but every model fell to between 12% and 21% on children, and ChatGPT dropped from 81.6% on large pneumothoraces to 42.2% on small ones. On measuring liver metastases against a radiologist's reference, Gemini reached an agreement score of 0.81, GPT-o3 0.52 and Claude 4 Opus 0.07.

Claude's pattern is consistent: it is the weakest of the majors at reading pixels and the strongest at reading radiology text. On structuring 3,949 head CT reports, Claude was significantly more accurate than GPT and Gemini for intracranial haemorrhage, and on 56 JAMA neuroradiology cases Claude 3.5 achieved the highest accuracy (80.4%) when given the image and the report together.

Two caveats. Every study above tested general models one generation or more behind the current ones — GPT-6 Astra, Claude Opus 5.5, Claude Fable 5.1 and Gemini 3.8 Flash have not yet been run through RadLE. And the RadLE cases are deliberately hard: on routine chest X-rays the chatbots do markedly better, as the pneumothorax study shows, and MedGemma does better still.

Speed

Two different things get called speed. Tokens per second is how fast text appears on screen; time per task is how long an agent takes to finish a job, which depends as much on how many tokens a model needs as on how fast it produces them.

If you are running agents, token efficiency is the number that sets your bill. Astra is not the fastest per token, but it needs about a third of the output tokens Grok 4.7 uses per task, and under a quarter of Opus 5.5's at max effort, which is why it has the lowest cost per task at its level on the index. Opus 5.5 narrows that gap at lower effort settings: Artificial Analysis finds four of its five effort levels on the best-value frontier among models scoring 50 or more.

Hallucination and accuracy

AA-Omniscience asks 6,000 hard factual questions and tracks two things separately: how often the model is right, and how often it makes something up when it does not know. A model can post a low hallucination rate simply by refusing to answer, so read the two columns together.

On the combined AA-Omniscience Index, which rewards right answers and penalises invented ones, GPT-6 Astra (44, at high effort) sat one point ahead of Fable 5.1 (43) before Opus 5.5 arrived. Artificial Analysis now lists Claude Opus 5.5 as the leader on AA-Omniscience, but its separate accuracy and hallucination figures were not in the launch summary. The very lowest hallucination rates on the board belong to small models that decline most questions. That is not a model you want; it is a model that has learned to say "I don't know". Also worth knowing: reasoning modes hallucinate two to three times more than plain modes on summarisation tests. If the facts matter, turn thinking off or give the model a document to work from.

Refusals and over-blocking

This is the category businesses ask about most and the one with the least data. No leaderboard measures "refused a perfectly reasonable request" across vendors, so the table below collects what has actually been published.

A model that refuses to look at malware is no use to the people whose job is looking at malware.

Image generation

Image models are ranked by blind human votes in the Artificial Analysis arena: two images from the same prompt, pick the better one, thousands of times. Elo is the native unit; higher is better and 30 points is a meaningful gap. OpenAI's September release, GPT Image 2.5, now holds the top two places both for creating images and for editing them.

Prices are Artificial Analysis's figure for 1,000 images at 1024×1024 on each maker's API. For diagrams, technical illustrations and anything with words in it, no benchmark measures "diagram quality" directly; text-rendering ability is the proxy, and GPT Image 2 and 2.5, Ideogram 4.0 and Nano Banana Pro lead it. For architecture diagrams you will usually get a better result asking a language model for SVG or Mermaid than asking an image model to draw one.

Video generation

Video models are ranked by blind votes in the Artificial Analysis Video Arena, and the board has turned over since the start of September. Google's Gemini Omni Flash now leads for clips with sound; the rest of the top five are built on Chinese models from Alibaba, MiniMax and ByteDance. Kling, which topped several leaderboards earlier in the year, has slipped to twelfth.

Elo scores are for clips with audio, as of 22 September. Without audio the order changes: Wan 3.0 leads text-to-video (1336) and Gemini Omni Flash leads image-to-video (1369). Prices are Artificial Analysis's figure for one minute of video on the maker's own API, except Seedance 2.5, which uses Higgsfield's credit price.

Seedance and Kling both come from Chinese companies, so prompts and uploads may fall under Chinese law, just as material sent to a US provider falls under US law. For client footage or unreleased products, check which company processes the data, and where, before you upload anything.

What people actually run

Everything above ranks models by how they score. This ranks them by how much work they are actually given. OpenRouter is a neutral exchange that routes requests across several hundred models from every major lab and publishes its own token volumes, so it is the closest thing to a usage leaderboard that exists. Here is where the tokens went in the seven days to 20 September 2026.

Volumes are prompt plus completion tokens for the seven days to 20 September 2026, so they predate the MiMo-V2.6 and Claude Opus 5.5 launches. OpenRouter lists variants of the same model separately, which is why DeepSeek V4 Flash appears twice. Source: OpenRouter (openrouter.ai/rankings), licensed under CC BY 4.0.

Read that against every ranking above and the mismatch is close to total. Only one model this page names as a category leader appears: DeepSeek V4.1 Flash, top of the CyberGym vulnerability-discovery test, which went from seventh place to first within ten days of its release. Claude Fable 5.1, GPT-6 Astra and Claude Opus 5 are nowhere near the top ten. Eight of the ten entries are Chinese; the two American ones are OpenAI's budget tier and a free Nvidia model. Counted by requests rather than tokens, the week beginning 14 September split DeepSeek 25.4%, Google 18.6%, OpenAI 17.0% and Z.ai 9.4%, with Anthropic on 2.7%.

Benchmarks measure the ceiling. Usage measures the floor, and almost all the work happens on the floor.

The pattern behind it is straightforward once stated. Almost every model in the top ten is open-weight, free or priced near the floor, and an exchange like this accumulates whatever is cheapest to serve. The frontier labs' pricing keeps their best models off it by design. That is worth knowing when you are choosing a stack: the models that win the tables above are bought for the hardest tenth of the work, and something like this list is what handles the rest.

What the work actually is

OpenRouter also classifies requests by task and reports the share of spend each takes. In the week to 12 September the split was a useful corrective to a page organised around medical, legal and scientific benchmarks.

Agent work is already more than a quarter of what people pay for, on a par with writing code. None of the professional benchmarks on this page measure it, which is why the tool-use and computer-use sections matter more than their position in the page suggests. The categories this page covers in most depth, medical and legal, do not appear in the spend breakdown at all.

One reason any of this can be wrong

For two weeks in August 2026 an anonymous listing called Ox Alpha carried 27 trillion tokens through the exchange, 12% to 14% of all traffic, with no lab attached to it. Labs routinely run unreleased models under codenames for a week or two before launch; Z.ai later confirmed that Ox Alpha was GLM-5.3-Flash and published the weights. So a usage table like the one above can be materially wrong about who is winning, for weeks at a time, while being exactly right about the totals.

Why everyone is talking about Kimi

Kimi is the assistant and model family from Moonshot AI, a Beijing start-up founded in 2023. Three things have put it in the headlines since July: a model that rivals the US frontier and can be downloaded, a business growing unusually fast, and a public accusation from a rival.

A frontier-class model you can download

Moonshot released Kimi K3 on 16 July and published the full weights on 27 July: 2.8 trillion parameters with 104 billion active per token, a one-million-token context window and native vision. At launch it took first place in the Frontend Code Arena, ahead of Claude Fable 5 and GPT-5.6 Sol, although Moonshot itself conceded that K3 still trailed those two models overall. The release landed in Washington as a question about how quickly a Chinese start-up had closed the gap. A US congressional committee had been pressing American companies over their use of Kimi since April, and in March Cursor admitted that its Composer 2 coding model was built on an earlier Kimi release.

It is not a bargain model. Moonshot charges $3 per million input tokens and $15 per million output, five times the input price of its predecessor, K2.6, although K3 is free to use in the Kimi app. The weights come under Moonshot's own licence rather than MIT: free for most businesses, but model-hosting companies with more than $20 million in annual revenue need a separate agreement.

A fast-growing business

Bloomberg reports that Moonshot is aiming for $2 billion in annualised revenue by the end of the year, double its August run rate, and OpenRouter shows K3 models generating up to 300 billion tokens a day. On 17 September Moonshot launched Kimi for financial services, wired into data from S&P Global Market Intelligence, Crunchbase, the US SEC's EDGAR filings, the IMF, the World Bank and the Federal Reserve's FRED database. It has reportedly filed confidentially for a Hong Kong listing.

The allegation

On 10 September Anthropic, which makes Claude, published a threat-intelligence report accusing Moonshot of more than copying. It alleged that in one ten-day window Moonshot relayed nearly 300,000 Kimi customer requests to Claude, mostly to Claude Opus, through 5,380 fraudulent accounts that appeared to be in Singapore and Japan; showed Claude's answers to users as if Kimi had written them; and kept the exchanges to train its own models, as part of a campaign in which Anthropic says more than 23 million responses were collected. The same report made distillation allegations against DeepSeek and MiniMax.

What it means if you are weighing Kimi up

• Know which model answers you. Whatever the truth of this case, a model name on screen is a claim, not a guarantee. Ask any AI vendor or reseller which model and which country process your data, and get the answer in the contract.
• Hosted Kimi is a Chinese service. Prompts sent to kimi.com or Moonshot's API are handled by a Chinese company. Keep client data off it unless your data-protection impact assessment clears it.
• The weights change the picture. Because K3 can be downloaded, it can run on infrastructure you control, and US hosts such as Together AI have committed to serving it from US infrastructure. At 2.8 trillion parameters, though, self-hosting is a job for a GPU cluster, not a single server; DeepSeek V4.1 Flash, at 552 billion parameters, is far easier to run.
• Judge it on price, not hype. At $3 / $15, K3 costs more per token than Grok 4.7 or Muse Spark 1.3 and scores lower than both. Xiaomi's MiMo-V2.6-Pro, also open-weight and under the more permissive MIT licence, now scores higher at $0.435 / $0.87. On score and price alone, K3 is hard to justify.

What it costs to use them yourself

API prices are in the model table above. For individual chat subscriptions, the whole market has settled on a $20-a-month standard tier, with budget tiers below it and power-user tiers of $100 to $300 above it. Prices are the vendors' US list prices; UK checkouts are billed in sterling with VAT added, so check the vendor page for the exact figure on the day.

Two things to check before relying on a consumer plan for work. Training terms differ: xAI's individual Grok plans may use your conversations for training while Grok Business ($30 per seat) does not, and Meta's cheapest Muse Spark API tier is cheap precisely because Meta may train on what you send. Availability differs too: some Gemini app features have excluded the UK and EEA at launch, and Claude Mythos 5.1 is not sold on any public plan. Business and Enterprise tiers from every vendor add admin controls, data-processing terms and, in most cases, EU or UK data residency.

Running an AI agent on your own PC

Everything above runs in someone else's data centre. A local agent keeps prompts and files on your own machine, in exchange for a much smaller model. It takes two pieces: a model runner such as Ollama or LM Studio, and an agent layer that plans and uses tools around the model. OpenClaw suits a persistent personal assistant, OpenHands and Cline suit software engineering, and goose suits general local automation. Hermes Agent, from Nous Research, learns reusable skills from each completed task and ships tool-call parsers for local models, so a small Qwen behaves predictably.

Memory decides almost everything. A 4-bit (Q4) model needs roughly 0.6 GB of video memory per billion parameters, before you add room for the conversation. The newest mixture-of-experts models help here: they hold many parameters in memory but only work a few billion of them per word, so they run far faster than their size suggests.

What tokens per second actually feels like

Models write in tokens, not words. In English prose a token is about three-quarters of a word, so 100 tokens is roughly 75 words; code uses more tokens per line because of symbols and indentation. An average adult reads silently at about 238 words a minute, which is only around 5 tokens per second. Anything above 10 or 15 tokens per second therefore outruns your eyes on a normal answer. Speed starts to matter when the model writes far more than you read: reasoning models think privately before answering, and coding agents write whole files and retry until the tests pass. Drag the slider or pick a machine to see the difference.

With that in mind, here is what each build can realistically run. Every PC row assumes the same base machine: an Intel Core Ultra 7 with 64 GB of RAM. The coding level column uses our own four-step scale, explained after the tables.

The Coding vs Opus 5 column shows how close the best coding model each machine runs at a usable speed comes to Claude Opus 5, on one shared yardstick: the Artificial Analysis Coding Index, where Opus 5 scores 78.0. The best coder that fits on a single 24 GB graphics card, Qwen 3.8 27B, scores 68.1, or 87% of Opus 5. The best that any machine on this page holds, Qwen3.8-Flash-Next at about 123 GB, scores 73.0, or 94%. Claude Opus 5.5 was released today and is not on the index yet; it beats Opus 5 on every coding test Anthropic published, and by 11 points on the independent Terminal-Bench 4.0 run, so the gap to Opus 5.5 is wider than these bars show.

Apple Mac Studio and MacBook Pro

Macs work differently. The processor and graphics share one pool of unified memory, so a Mac can load models far larger than any consumer graphics card holds. The trade-off is speed: writing each token is limited by memory bandwidth, where Nvidia cards are faster, while reading your prompt is limited by raw compute, where the M5 generation made its biggest jump. In Apple's own measurements, M5 produced the first token 3.3 to 4 times faster than M4 but subsequent tokens only 19–27% faster. For coding agents, which send long prompts full of code, that first-token speed is the one you feel.

Two supply notes. There has never been an M3 Max Mac Studio: that chip shipped only in the MacBook Pro, so it is listed here as a laptop. And the memory shortage has hit Macs hard: Apple removed the 512 GB M3 Ultra option in March 2026, and by May the M3 Ultra Mac Studio was sold only with 96 GB. The M5 Max and M5 Ultra Mac Studio, announced on 26 August, restore the 512 GB tier.

M5 Ultra prices start at $5,499 with 96 GB and run to $18,299 fully loaded; the 96 and 256 GB models ship from 22 September and the 512 GB model in late October. M5 Ultra speeds are published estimates, not measurements, until shipped units are benchmarked. Measured figures depend on the runtime (MLX, Ollama, llama.cpp), quantisation and context length, so treat them as a guide.

Dedicated AI computers

A new class of machine sits between the gaming PC and the datacentre. Some copy Apple's approach of one large pool of shared memory, but use AI chips instead of desktop processors. Others stack professional graphics cards with 96 GB each. All of them run the same open-weight models as the builds above. What changes is how big a model fits, how fast it runs, and how many people it can serve at once.

Every machine here tops out at coding level 3. Even the $85,000 GB300 stations run the best open model, which scores 46 on the intelligence index, against 58 for Claude Opus 5.5 in the cloud. What the expensive machines buy is capacity and speed for a whole team, with nothing leaving the building, rather than frontier intelligence. Bandwidth decides single-user speed: the GB10 and Ryzen AI Max boxes share the same 273 GB/s, about a third of an M3 Ultra, so they favour mixture-of-experts models over dense ones.

What each option costs

Prices are US list or street prices seen in August and September 2026, or our own estimates where no price has been published. Memory and graphics card shortages have pushed almost every figure up this year, so check the day's price before you buy. UK checkouts are in sterling with VAT added. Electricity, software and support are not included. The two new columns show the ceiling of each option: the strongest model it can hold, with its Artificial Analysis Intelligence Index score where one exists, the coding level from our four-step scale, and the fastest published speed. Speeds are for different models, so compare them with care: a fast figure on a small model is not the same as a fast figure on a large one. All PC rows share the Core Ultra 7, 64 GB base. The Coding vs Opus 5 column is explained above the first local table.

For scale: $4,699, the price of a DGX Spark, would pay for a $20-a-month Claude or ChatGPT plan for about 19½ years, or buy about 235 million Claude Opus 5.5 output tokens at $20 per million. Local hardware pays back through privacy, predictable cost and very high volume, not by being cheaper for occasional use.

The best setup for your budget

Pulling the tables together, here is what we would buy at each price point today, and what is worth waiting for. Prices are US prices in September 2026. "A PC you already own" assumes a desktop with a free graphics slot and a power supply big enough for the card. The Coding vs Opus 5 bar is the same yardstick used in the tables above.

The curve flattens fast. $1,000–2,000 already reaches 87% of Opus 5 on the Coding Index, and a further $9,000 adds only another seven points. What bigger budgets mainly buy is speed, room for larger general-knowledge models and capacity for a team. No budget here reaches Claude Opus 5.5, which a $20-a-month plan includes.

Local against the cloud: speed

The fastest cloud models write several times faster than any local machine, and they do it while running much larger models. The local figures below are for the best model each machine runs comfortably, so they are not like-for-like: a local 27B model is smaller than anything in the cloud rows.

Tokens per second, single user. Bars are scaled to Gemini 3.8 Flash. Cloud figures are Artificial Analysis output-speed measurements; Claude Opus 5.5 has not yet been measured independently. A local machine serves one person at full speed; a cloud API serves you at this speed however many of your staff use it at once.

Local against the cloud: intelligence and coding

This is where the gap is widest. On the Artificial Analysis Intelligence Index, the best model a machine under $20,000 can hold is DeepSeek V4.1 Flash, on a 512 GB M5 Ultra. It scores 39, level with the budget cloud tier. Only the $85,000-plus GB300 stations can hold the strongest open model, MiMo-V2.6-Pro, and no local machine can run the frontier models at all.

Artificial Analysis Intelligence Index v4.3. Qwen 3.8 27B, the strongest model that fits on a single 24 GB graphics card, scores 34 at its highest effort setting; its predecessor Qwen 3.6 27B scores 21. On coding the picture is more mixed. On the Artificial Analysis Coding Index, Qwen 3.8 27B scores 68.1 against 78.0 for Claude Opus 5, and on the older SWE-bench Verified test Qwen 3.6 27B's 77.2% is within four points of Claude Opus 4.6's 80.8%. On the harder Terminal-Bench 4.0, the best open model, MiMo-V2.6-Pro, reports 34.9% against 59.6% for Claude Opus 5.5.

Desktop coworkers compared: Claude, ChatGPT and Copilot

All three vendors now sell an assistant that works across your files, apps and desktop rather than just answering questions. Anthropic merged Claude's chat and Cowork modes into one window on 16 September. OpenAI rebuilt its desktop app around ChatGPT Work on 9 July, combining Chat, Work and Codex. Microsoft's Copilot Actions carries out tasks in a separate agent workspace on Windows, but as of September 2026 it has not reached general availability and remains an experimental preview.

Connector counts come from Node8's August 2026 capture of both public directories and change weekly. OSWorld scores are each vendor's own and use different setups. Timings are from The New Stack's three-task comparison published on 22 September.

Sixty common business apps and where they connect

Most business software now publishes an MCP server, the plug that lets an AI assistant read and act inside it. What differs is whether the app is a one-click listing or something your IT team has to wire in. These are sixty of the most widely used apps with MCP servers, not a ranking by usage.

Claude and ChatGPT status comes from Node8's August 2026 capture of both directories. Gmail, Google Drive and Google Calendar in ChatGPT are OpenAI's own built-in apps; Microsoft 365 appears there as four separate apps (Outlook Email, Outlook Calendar, Teams and SharePoint). The Copilot column refers to Microsoft 365 Copilot and Copilot Studio, where an administrator adds third-party MCP servers as tools. Copilot Actions on Windows cannot use any of these.

What this means for a UK business

• Day-to-day development: Claude Opus 5.5 is now the best capability per pound for PHP, C#, Python and SQL work, at $4 / $20 per million tokens, with Sonnet 5 at $2 / $10 for routine tasks. Keep GPT-6 Astra as a second opinion on the hardest long-running jobs, and Gemini 3.8 Flash for bulk, low-stakes tasks, but budget for its price doubling on 1 January.
• Security work: cloud models over-block, and Opus 5.5 is stricter on security requests than Opus 5. Use Opus 5 where the cloud is acceptable, look at Anthropic's Cyber Verification Program if this is routine work for you, and use a self-hosted DeepSeek V4.1 Flash or GLM-5.3 for malware triage and vulnerability hunting, which also keeps client code inside the UK.
• Client data: prefer Business or Enterprise tiers with UK/EU data-processing terms, or open weights on your own hardware. A $20 consumer plan is not a data-processing agreement.
• Desktop agents: Claude and ChatGPT can now operate your PC and your Microsoft 365 apps; since July, Claude’s Microsoft 365 connector can send mail, manage calendars and create files in OneDrive and SharePoint. Pilot on a test machine with backups, grant access app by app, and keep Copilot Actions to a test device until Microsoft releases it properly.
• Know who is actually answering: the Kimi allegation, and a fraudulent "discount Claude" reseller described in the same Anthropic report, both make the point that a model name on screen is a claim. Buy AI access through authorised channels, ask which model and which country process your data, and get it in writing.
• High-volume, low-stakes work: this is where the usage table above should change your shortlist. If you are processing thousands of documents rather than solving hard problems, the cheap open-weight models carrying most of the world's tokens, such as DeepSeek V4.1 Flash, GLM-5.3-Flash and Tencent's Hy4, will do it at a fraction of frontier pricing, and Xiaomi's new MiMo-V2.6-Pro is worth adding to that shortlist.
• Anything that leaves the building: legal, medical and client-facing text gets a human read every time, whichever model wrote it. Fable 5.1 scored 67% on the hardest published factual test, so even a top model still gets about a third of those questions wrong or unanswered.
• Marketing assets: GPT Image 2.5 for graphics with words in them, or Ideogram 4.0 if you want open weights. For video, Google's Gemini Omni Flash or Veo 3.1 are the simplest to license, Seedance 2.5 through Higgsfield suits 30-second multi-shot adverts, and fal's MiniMax H3 Max is the pick if cost per minute matters most. Whatever you use, keep real people, brands and film characters out of the prompt.
• Keep a fallback: models disappear. The Sora 2 API closes on 24 September, and in June Anthropic suspended Fable 5 for nearly three weeks to comply with US export controls, restoring access on 1 July. Build so that switching model is a configuration change, not a rewrite.
• Expect this page to change: Claude Opus 5.5 arrived on 22 September with Sonnet 5.5 and Haiku 5.5 due in the coming weeks, StepFun says Step 5 weights will follow on 15 October, Google's Gemini 3.5 Pro is still in partner testing, DeepSeek has said a V4.1 Pro is coming, and a Kling 4.0 is widely expected. We will update the figures as they move.