The AI model comparison table: every model that matters, five numbers each

In the past twelve days: xAI shipped Grok 4.5, OpenAI shipped three models at once, and Moonshot shipped a 2.8-trillion-parameter one with open weights promised within the week. Any “best AI models” article written in June is already wrong in at least four rows.

So this page is not an article. It is a table: every model that matters right now, five numbers each (price in, price out, context, license, benchmark standing), sortable and filterable, with a date stamp and a source for every figure. Below it, the short course in reading any model table without being fooled. Bookmark the table; the reading skills are the part that won’t go stale.

The leaderboard turned over twice this month

Consider what July 2026 alone did to any static comparison. On July 8, xAI released Grok 4.5 at $2 per million input tokens. On July 9, OpenAI replaced its flagship with three models named Sol, Terra, and Luna, spanning a 5x price range by themselves. On July 16, Moonshot released Kimi K3: 2.8 trillion parameters, a 1M-token context window, and a claim to the largest open-weight model ever shipped, with the weights themselves due July 27.

Three releases, twelve days. Each one reshuffled at least one column that buying decisions hang on. And the quieter changes matter as much as the launches: Anthropic has its mid-tier Sonnet 5 on an introductory $2 input price that expires August 31, Meta’s newest model dropped out of the comparison entirely by refusing to sell API access, and the “largest open model ever” is, for one more week, a model card without a download link.

That pace is why this page is built as a maintained reference instead of a listicle. The rows get edited as models ship, the “updated” stamp moves, and numbers that stop being true get corrected rather than left to rot. A comparison you can’t date is a comparison you can’t trust.

The table: sort it, filter it, argue with it

A note on what made the cut. The table covers language models you can buy through an API or download and run, because those share comparable units. Image, video, and audio models price by the picture and the second instead of the token, so forcing them into these columns would produce nonsense; they get their own comparisons, like our video model matchup.

Eighteen rows, six columns. Prices are first-party API list prices in dollars per million tokens. Context is the advertised window. “Index” is the Artificial Analysis Intelligence Index, a composite of ten evaluations and the closest thing the industry has to one comparable number. A dash means no published score, not a zero.

If you’re new to the units: models read and write in tokens, chunks of text averaging about three quarters of a word. You pay one rate for what the model reads (input) and a higher rate for what it writes (output), because generating text costs the vendor far more compute than reading it. The gap is standardized enough that OpenAI’s whole GPT-5.6 family prices output at exactly 6x input. For most everyday use, input dominates the bill; for long reports and code generation, output does.

Two reading tips before the caveats. First, filter by Open weights and notice that the open column now starts at an Index score of 57: frontier territory, not a consolation bracket. Second, sort by Cheapest input and notice that the bottom half of the table is almost entirely open models. Those two sorts, side by side, are the 2026 market in one glance.

Two good models can sit 70x apart on price

The price column has the widest spread in the table. Claude Fable 5, the current Intelligence Index leader, lists at $10 per million input tokens and $50 per million output. DeepSeek V4-Flash lists at $0.14 and $0.28. That is 71x on input and 179x on output between two models that both handle everyday work competently. Even staying inside the frontier tier, Kimi K3 scores within three Index points of Fable 5 at less than a third of its price.

Make it concrete. Summarizing a 10,000-word report (roughly 13,000 tokens in, 1,000 tokens out) costs about 18 cents on Fable 5 and about a fifth of a cent on V4-Flash. Run that summary daily and one choice costs $66 a year while the other costs pennies. The expensive model is often worth it, but it is worth it for the hard 10% of your work, not for the routine 90% that a $0.14 model handles identically.

The sticker is also not what you pay. Every vendor in the table now sells the same three discounts under different names: cached input (repeated prompt prefixes billed at roughly a tenth of list), batch processing (about half price for answers that can wait), and off-flagship tiers (Sonnet 5 is even on an introductory $2/$10 sale through August 31). Anthropic’s pricing page adds a subtler trap: its newest models use a tokenizer that produces roughly 30% more tokens for the same text, so a per-token price cut can be smaller than it looks once the meter counts differently.

That is why the honest unit is never the token. A model that costs 3x more per token but solves the task in one attempt beats a cheap model that needs five tries. We’ve run that math across providers in cost-per-task, not cost-per-token; the table’s price columns are the start of a cost estimate, not the end.

Context windows are advertised, not guaranteed

Sort by context and the column looks almost uniform: 1M tokens has become the standard badge, from Fable 5 down to DeepSeek’s $0.14 model. For scale, a million tokens is roughly 750,000 words, which is the length of eight novels fed into a single question. Two years ago that capacity was a headline feature on one model; today it ships on a model that costs 14 cents. The badge hides two asterisks, though.

The first is performance. Advertised context is how much text the model accepts; effective context is how much it can use before recall degrades, and on benchmarks built to measure the gap the effective number is routinely half the sticker or less. That gap is the subject of our long-context deep-dive, and it is why the table treats context as capacity, not competence.

The second is price cliffs. Several vendors charge more per token once a prompt crosses a threshold: Grok 4.5 doubles to $4/$12 at 200K prompt tokens, Gemini 3.1 Pro doubles input above the same line, and GPT-5.6 bills 2x input past 272K. Anthropic went the other way this generation and charges flat rates across the full 1M window. Same column in the table, four different bills for the same long document.

“Open” is five licenses that don’t agree

The weights column splits the table into open and closed, but “open” is doing a lot of work in that sentence. The open rows in this table span four genuinely different legal regimes, and picking between them matters more than most spec-sheet differences the moment a model touches a product. A benchmark point is a rounding error next to a license clause that says your use case needs a separate agreement.

The direction of travel is worth a row of its own in your head. In April, Google moved the core Gemma family from its custom terms to plain Apache 2.0, the permissive standard. The same month, Meta went the opposite way, shipping its first closed frontier model. Licenses are not settled infrastructure; they are strategy, and they change with it.

Two rows deserve a closer look. Kimi K3 is listed as open with an asterisk because, as of this writing, the weights are a promise with a date attached, not a download; a press release is not a license until the files land. And Muse Spark sits at the bottom of the table as the strangest row of 2026: Meta, the company that created the open-weights era, shipped its first fully closed frontier model in April and walked away from the Llama line. The open frontier is now carried almost entirely by Chinese labs and a handful of startups.

The practical read: if the weights are MIT or Apache, you can build on them with ordinary open-source hygiene. Anything else, read the actual file before shipping. The distance between “MIT” and “modified MIT” is one clause, and that clause is about you.

A dated number beats an undated one

The Index column deserves its own warning label, because benchmark numbers are the most quoted and least understood cells in any model table.

First, configuration moves scores. GPT-5.6 Sol scores 59 on the Intelligence Index at maximum reasoning effort, 58 one notch down, and 56 at the setting most apps actually run. A single model is a range, not a point, and vendors quote the top of their range. The table pins every score to the model’s best configuration so at least the comparison is consistent.

Second, benchmarks disagree because they measure different jobs. The Intelligence Index averages exam-style evaluations. LMArena counts millions of blind human votes on which answer people prefer, which is why Claude Opus 4.8 tops it while carrying no Index score in this table. SWE-bench Verified measures something narrower and harder: fixing real bugs in real repositories. Coding tables reorder everything: DeepSeek V4-Pro’s Max configuration posts 80.6% on SWE-bench Verified while sitting at 44 on the general Index. Neither number is wrong. They answer different questions, which is why the table names its benchmark instead of inventing a mystery five-star rating.

Third, every number here has a date, and that is a feature. A score without a date is folklore. If you take one habit away from this page, take that one, and then go one step further: the benchmark that actually settles your model choice is the one you write yourself from twenty examples of your own work.

Keep three rows, recheck monthly

Eighteen rows is for comparing. For working, you need three: a default (the best model you can afford for work that matters), a workhorse (a cheap tier or open model for volume), and a local fallback (an open model that your machine can actually run, for private files and offline days). Pick them from the table today, and the eighteen rows collapse into a decision you only revisit when something changes.

A reasonable trio as of this stamp: Sonnet 5 as the default while the introductory price holds, DeepSeek V4-Flash as the workhorse at $0.14, and Qwen 3.5 9B as the local fallback if your machine has 16 GB of memory. Yours will differ with your work; coders should look hard at the Ornith and Laguna rows, and anyone drafting long documents should weigh output price over input. The point is not this trio. The point is that three deliberate rows beat eighteen half-remembered ones.

Knowing when something changed is what this page is for. Come back after the next launch day, check the updated stamp, re-run your two favorite sorts, and see whether your three rows still hold. Most months they will. The months they don’t are exactly the months this table earns its bookmark.

Disclaimer: The license descriptions in this post are plain-English summaries, not legal advice. Model licenses change between releases, and what counts as permitted use depends on your situation. Read the license text that ships with the model, and consult counsel before relying on it in a product. Details reflect sources available as of July 2026.