Best AI Coding Assistant 2026: 10 Tools, 60 Questions, 15 Answers

Last updated: August 21, 2026

An AI coding assistant reads your repository and writes or reviews code on your behalf. Ten of them are covered here. They are not ten versions of one purchase: some sell a seat, one sells a flat team plan, one is bundled inside a wider subscription, one publishes a free individual route with no price attached, and one appears in this briefing only through model-layer metering.

One-line finding: put six buyer questions to ten products, and the pages read here answer fifteen of the sixty. No product answers more than three.

Five ways these products are sold

Before any figure can be compared, the objects have to be separated. These ten are sold under five different access models, and a ratio calculated inside one of them does not transfer to another.

Even the unit differs. GitHub Copilot denominates in AI credits and states that one credit equals one US cent. Kiro denominates in credits and defines a credit as a unit of work metered to two decimal places. Cursor runs two separate pools, one for its own models and one for third-party models charged at the model’s API price. Command Code denominates in dollars of credits spent at the underlying model rate. Augment denominates in dollars measured across model inference, a service fee, and compute time. Codex denominates in credits per million tokens, and separately in local messages and cloud chats inside a rolling five-hour window. Grok Build is priced per million tokens.

Three publish no size at all on the pages read here. Claude’s individual comparison table shows a context window and relative multipliers but no allowance figure. Devin states that each paid plan carries an allowance refreshing daily and weekly, without printing its size. Antigravity’s individual card states basic weekly rate limits without a number.

Sources: GitHub, Plans and pricing, read 21 August 2026 · AWS, Kiro Pricing, read 21 August 2026 · Anysphere, Cursor Models and Pricing, read 21 August 2026 · Command Code, Pricing, read 21 August 2026 · Augment Code, Pricing, read 21 August 2026 · OpenAI, Codex Pricing, resolves to a ChatGPT Learn page, read 21 August 2026 · Anthropic, Plans and Pricing, read 21 August 2026 · Cognition, Devin Plans and Pricing, read 21 August 2026 · Google, Antigravity Pricing, read 21 August 2026 · SpaceXAI Docs, Pricing, last updated 3 July 2026

What the published record answers

Six questions were put to all ten products. The counting rule is published in Methodology and is designed so that anyone can rerun it and disagree with the result.

DERIVED. Cells marked stated: 2 + 1 + 1 + 1 + 1 + 2 + 1 + 2 + 3 + 1 = 15. Total cells: 10 products × 6 questions = 60. Share stated: 15 ÷ 60 = 25.0 percent. Cells marked partial: 7. Cells where an element was present but its value could not be read: 3. Cells with no relevant statement found: 35. Check: 15 + 7 + 3 + 35 = 60. Counting stated and partial together: 22 ÷ 60 = 36.7 percent. Highest per product: 3 of 6, which is 3 ÷ 6 = 50.0 percent, reached by one product.

Two columns carry almost nothing. Question B is answered by one page out of ten, and that answer is disputed by another part of the same page, as set out in section 03. Question C is answered by two, and by a third only in part.

The matrix is bounded by what was read. Six of the ten products contributed one page each, so a blank cell means the question was not answered on the page or pages opened for this briefing. It is not a finding that the vendor is silent everywhere. Methodology publishes the page count per product so that the bound is visible.

Sources: the eighteen pages listed in section 13, all read 21 August 2026. Each cell is supported by the section that discusses it.

Indemnity and training, tier by tier

The two rights that matter most in a procurement review are whether the vendor trains on your code and whether it will stand behind the output if a copyright claim arrives. Only one of the ten pages read here addresses the second question at all, and that page does not appear to agree with itself.

GitHub Copilot. The plan comparison table carries a section headed Policies and management. The row for IP indemnity displays Not included in all four individual columns: Free, Pro, Pro Plus, and Max. The rows for enterprise-grade security, user management in github.com, usage metrics, and SAML SSO authentication display Not included in the same four columns. Switching the same page to its business view, IP indemnity together with data privacy appears in the listed benefits of the Business plan at 19 US dollars per user per month.

The same page also carries collapsed FAQ entries under Responsible AI and under Plans and pricing that address intellectual property and the differences between the individual and organizational offerings. Those entries were not expanded in the reading recorded here. Whether they narrow the table row, restate it, or contradict it is unresolved.

This briefing therefore does not state that an individual Copilot subscriber lacks indemnity. It states what the table row displays, records that other surfaces on the same page were not read, and puts the question at the top of its recheck list.

Training terms. Five of the ten pages state one. Anthropic’s individual comparison table shows Model training with the value Opt-out in all four individual columns. Google’s Antigravity additional terms state that Google records and stores user data, interaction data, related metadata, and feedback, collectively defined as Interactions; that Google uses Interactions to evaluate, develop, and improve Google and Alphabet research, products, services and machine learning technologies; that Google employees and contractors may access, view, review and use Interactions; and that a settings screen changes the preference. The same document opens by stating that these terms do not apply to anyone accessing the service through Gemini Enterprise on Google Cloud, Gemini Enterprise for Business, a Google Workspace subscription on the Pre-GA Offering Terms, or a Gemini Enterprise Agent Platform API key. xAI’s security documentation states that it never trains on API inputs or outputs without explicit permission, and that requests and responses are stored encrypted at rest for thirty days for auditing and then deleted. Augment Code prints No AI training allowed on both its Business and Enterprise cards. Command Code’s comparison table carries a row reading No training on your code across every plan from the one dollar tier upward.

DERIVED. Pages stating a training term: 5 of 10, which is 5 ÷ 10 = 50.0 percent. Two further pages carry a training row whose value could not be read reliably.

Sources: GitHub, Plans and pricing, Policies and management section and business view, read 21 August 2026 · Anthropic, Plans and Pricing, Models and usage section, read 21 August 2026 · Google, Antigravity Additional Terms of Service, opening paragraph and clauses 3 and 5, read 21 August 2026 · SpaceXAI Docs, FAQ Security, last updated 27 July 2026 · Augment Code, Pricing, read 21 August 2026 · Command Code, Pricing, Privacy and security section, read 21 August 2026

Protection that changes what you get

Where a data protection control is documented, the same document usually states what changes when it is on. Two products publish that detail, and a third attaches the same shape to a compliance region.

xAI. The security documentation states that by default all API requests and responses are stored encrypted at rest for thirty days for auditing in the event of suspected abuse or misuse, and are deleted automatically after thirty days. It then describes Zero Data Retention and states that for most customers it does not recommend enabling it, because capabilities that depend on stored data become unavailable. The table of unavailable features lists per-API-key request logging, the stateful Responses API, the Files API, the Collections API used for retrieval, the Batch API, deferred completions, stored image and video outputs, and voice agent conversation history. DERIVED: eight table rows counted under the opened section. The control is applied at team level with no ability to enable it for specific API keys.

The same document separates the surfaces of one product. For the Grok Build command line interface it states that with Zero Data Retention enabled no trace or code data is retained, and that without it, privacy settings are available as a slash command and in settings so code data retention can be disabled separately. For Build mode on the grok.com website and the mobile apps it states that chats and build sessions are retained for product functionality, that unlike regular chat, private chats are not currently available for Build mode, and that if the improve the model toggle is enabled in data controls, Build mode chats and sessions including data from created apps may be used for product and model improvements.

Cursor. The models and pricing page states that opting in to regional data residency incurs a ten percent uplift on model pricing for eligible models, and refers the reader to a separate privacy and data governance page for supported regions, models, functions, and policy. That page was not read for this briefing. The eligible plans, eligible regions, activation method, model boundary, and any capability exclusions are therefore not stated here, and the ten percent figure should not be read as a description of a control available on every plan.

Kiro. The pricing FAQ states that pricing in AWS GovCloud regions is approximately twenty percent higher than standard commercial region pricing, that all four paid tiers are subject to that adjustment, that the free tier is not available in GovCloud, and that GovCloud requires enterprise authentication through AWS IAM Identity Center.

Antigravity. The route rather than a setting carries the change. The additional terms do not apply to anyone accessing the service through Gemini Enterprise on Google Cloud, so moving out of those terms means moving to a different purchase route.

Sources: SpaceXAI Docs, FAQ Security, last updated 27 July 2026 · Anysphere, Cursor Models and Pricing, Model pricing section, read 21 August 2026 · AWS, Kiro Pricing, Common Questions, read 21 August 2026 · Google, Antigravity Additional Terms of Service, opening paragraph, read 21 August 2026

How much of your allowance reaches the model

One vendor publishes the composition of its allowance. A second publishes a platform charge that survives a customer’s own API key. A third states that it adds no model markup while displaying a separate unquantified fee.

Augment Code. The Business plan is one hundred dollars per month with no per-seat charge, covering up to fifty seats, and includes one hundred dollars of usage every month. The pricing FAQ states that the usage is measured in dollars across model inference billed at the provider’s public API list price, a flat forty percent fee on model inference with no fee on compute, and Cosmos compute time.

DERIVED. For a month spent entirely on model inference with no compute consumed, L + 0.40L = 100, so 1.40L = 100, so L = 100 ÷ 1.40 = 71.43 dollars of inference at list price. The fee component is 100 − 71.43 = 28.57 dollars, which is 28.57 ÷ 100 = 28.57 percent of the included allowance. A flat forty percent fee on inference and a 28.57 percent share of the allowance describe the same published term from two directions. The page prints the first. A budget holder needs the second.

Cursor. On Teams and Enterprise plans, third-party model requests carry a Cursor Token Rate of 0.25 US dollars per million tokens on top of the model’s own API pricing. The page states that it applies to included usage, to on-demand usage, and to bring-your-own-key usage, and that Auto Cost and all first-party Cursor models are exempt. Supplying your own key does not remove the platform charge. Selecting the vendor’s own models does.

Command Code. Three separate statements appear on the page and they are not opposites of one another, because they sit on different billing layers.

The relationship between an unquantified processing fee and a stated absence of model markup is unresolved on the page read here, and nothing in this briefing treats one as cancelling the other.

For the remaining seven products, no platform fee or allowance composition was stated on the pages opened. No fee should be inferred from that silence.

What can be compared, strictly within one product’s own price list, is the direction the allowance moves as the tiers climb.

Three directions across five price lists. Two rise, one falls, one is flat, and one is a single tier. None of the four directions is discoverable from a headline price.

Sources: Augment Code, Pricing, FAQ item 01, read 21 August 2026 · Anysphere, Cursor Models and Pricing, Cursor Token Rate section, read 21 August 2026 · Command Code, Pricing, plan cards, comparison table and FAQ, read 21 August 2026 · GitHub, Plans and pricing, GitHub AI Credits section, read 21 August 2026 · AWS, Kiro Pricing, read 21 August 2026

The portion the vendor labels variable

Some of what these pages present as a monthly allowance is described, on the same page, as changeable or shared. The vendors’ own words are used here, and no timing, notice period, or treatment of an active billing period is inferred from them.

GitHub Copilot splits AI credits into two lines. Base credits are 10 dollars on Pro, 39 dollars on Pro Plus, and 100 dollars on Max. The Flex allotment is 5 dollars, 31 dollars, and 100 dollars respectively, and the table describes it as variable additional usage on top of your base, adding that Flex allotments may change over time. A footnote on the same page repeats that the Flex quota is not fixed and may change in future.

DERIVED. Base credit value ÷ price: 10 ÷ 10, 39 ÷ 39, 100 ÷ 100, each equal to 1.00. Flex as a share of the displayed total: 5 ÷ 15 = 33.3 percent on Pro, 31 ÷ 70 = 44.3 percent on Pro Plus, 100 ÷ 200 = 50.0 percent on Max. The share of the advertised total that the page itself labels variable rises with the price of the plan. Both facts appear on the page. The page does not connect them, and neither does this briefing.

OpenAI Codex states a shared draw rather than a variable one. The pricing page states that ChatGPT Work and Codex share usage, and that ChatGPT Work usage inside ChatGPT uses the same pricing, credits, and usage limits as Codex. It further states that usage limits are shared with other agentic features once pricing for those features is effective, and that this currently includes ChatGPT for Excel on Plus and Pro. A budget bought for code is drawn on by work that is not code.

Grok Build. The pricing page states that when a request is deemed to be in violation of usage guidelines by the system, the generation is still charged, and that for violations caught before generation in the Responses API a fee of 0.05 dollars per request applies. It also states that the twenty percent Batch API discount applies to a named list of models and that models not on that list have no batch discount. Priority Processing is billed at twice standard rates across all token types. This briefing states only that a charge may occur under the published conditions, and draws no conclusion about the usefulness of the result.

Claude. The individual plan cards carry a general reservation: usage limits apply, and price and plans are subject to change at Anthropic’s discretion. This is a reservation about plans and prices. It is not a statement identifying a variable portion of an allowance, and it is not counted as one in section 02.

Antigravity. Asterisks appear against several plan items and a footnote reading that these are rolling out to select customers appears on the page. Which symbol maps to which cell was not established from the reading recorded here, so no item is described as variable on that basis.

Sources: GitHub, Plans and pricing, GitHub AI Credits section and footnote 3, read 21 August 2026 · OpenAI, Codex Pricing, read 21 August 2026 · SpaceXAI Docs, Pricing, Batch, Priority and Usage Guidelines sections, last updated 3 July 2026 · Anthropic, Plans and Pricing, read 21 August 2026 · Google, Antigravity Pricing, read 21 August 2026

Quota rises, parallelism does not

Two pages state a concurrency figure. On one of them the figure does not move with the price.

Devin. The plan comparison table lists a row for concurrent sessions limit. The Pro plan at 20 dollars per month displays Up to 10. The Max plan at 200 dollars per month displays Up to 10. Unlimited appears in the Team and Enterprise columns. DERIVED: the price ratio between those two plans is 200 ÷ 20 = 10.00, and the displayed concurrency ratio is 10 ÷ 10 = 1.00. The plan card describes what the higher price buys as significantly higher quotas. It does not describe more sessions at once.

Augment Code. The comparison table lists 50 concurrent sessions under Business and unlimited concurrent sessions under Enterprise, with standard size compute on Business and custom size and multi-region compute on Enterprise.

Four further products state a sharing or relative limit rather than a concurrency number. GitHub Copilot’s pooled usage row, described as included AI credits shared org-wide with admin controls for spending, displays Not included in all four individual columns, while the business view describes the Business plan as being for teams that want to share and manage credits across the team. Kiro’s FAQ states that for team usage each developer needs their own subscription. Cursor states that its Teams plan offers Standard seats at 40 dollars and Premium seats at 120 dollars per user per month, with Premium raising agent limits to five times Standard. Command Code’s table describes rate limits as Standard on the one, ten, and twenty dollar plans, Higher on the hundred dollar plan, and Highest on the two hundred dollar plan, while listing a context window of up to one million tokens on every plan including the one dollar tier.

Kiro’s unit economics say the same thing about upgrades. DERIVED: moving from Pro to Power multiplies credits by 10,000 ÷ 1,000 = 10.00 and multiplies price by 200 ÷ 20 = 10.00, so the price per included credit is unchanged at 0.02 dollars, and the add-on rate remains 0.04 dollars, which is 0.04 ÷ 0.02 = 2.00 times the included rate at every tier.

The model layer applies its own ceiling. xAI’s pricing page states that models with long context pricing bill the long context rates for all tokens in a request once its prompt reaches the model’s long context threshold, and the threshold shown is 200,000 tokens. DERIVED: comparing the short and long columns for each listed model, the long context rate is exactly 2.00 times the short context rate for input, cached input, and output. The portion of the advertised window that sits above the threshold is (256 − 200) ÷ 256 = 21.9 percent for a 256,000 token window, (500 − 200) ÷ 500 = 60.0 percent for a 500,000 token window, and (1,000 − 200) ÷ 1,000 = 80.0 percent for a one million token window. The larger the advertised context, the larger the share of it priced at double, and crossing the threshold reprices the whole request rather than the excess.

Sources: Cognition, Devin Plans and Pricing, Compare Plans table, read 21 August 2026 · Augment Code, Pricing, Compare plans table, read 21 August 2026 · GitHub, Plans and pricing, individual and business views, read 21 August 2026 · AWS, Kiro Pricing, Common Questions, read 21 August 2026 · Anysphere, Cursor Models and Pricing, Teams section, read 21 August 2026 · Anysphere, Cursor Changelog, entry dated 19 August 2026 · Command Code, Pricing, comparison table, read 21 August 2026 · SpaceXAI Docs, Pricing, Text API table, last updated 3 July 2026

Same plan, different rendered state

A pricing page is not a fixed document. What it displays depends on where the reader is, which country the account is set to, and which purchase channel was used. The figures below are labelled with the state they were read in, because without that label they are not checkable.

Anthropic, Japan-rendered state. Read from Japan on 21 August 2026, the individual cards displayed Pro at 18 dollars per month with an annual subscription discount, 220 dollars billed up front, and 22 dollars if billed monthly, with Max from 110 dollars per month. Below the cards the page displayed that prices include ten percent Japanese consumption tax. A reading from a different country or account state may display different figures, and this briefing makes no claim about what those would be.

DERIVED, applied only to the state described above. Removing the stated ten percent: 22 ÷ 1.10 = 20.00, 220 ÷ 1.10 = 200.00, and 110 ÷ 1.10 = 100.00. The card’s monthly figure against the annual total: 220 ÷ 12 = 18.33, which the card displays as 18. Comparing the two billing routes at the displayed figures: 22 × 12 = 264, and 264 − 220 = 44, so the annual route is 44 ÷ 264 = 16.7 percent lower over twelve months.

Kiro states the opposite convention. Its prices are exclusive of applicable taxes and duties including value added tax and sales tax, and it states separately that for customers with a Japanese billing address, use of Kiro is subject to Japanese consumption tax. It lists the countries in which a paid plan can be purchased with a billing address, notes that enterprise plans are purchased and billed through AWS so supported locations follow AWS regions instead, and states that model availability can vary by country or region because its model offerings align with each model provider’s usage and geographic requirements.

Cursor runs a plan that exists in one country. Start is described as available to developers in India at 649 rupees per month, tax inclusive, billed in Indian rupees, and the page states that every other individual plan displays its price before tax. On Start, all three first-party models run in non-fast mode, Grok 4.6 and Grok 4.5 are fixed at medium effort, and effort levels and fast mode cannot be enabled. Start excludes the third-party model pool, on-demand usage, Bugbot, Auto, automations, and the Cursor SDK.

OpenAI marks a set of Codex capabilities Limited in every plan column, including computer use in the browser, use with Chrome, record and replay on macOS, and memories, with a footnote stating that the feature is currently limited to only specific regions.

GitHub Copilot restricts by purchase channel rather than by country. A footnote on the plans page states that users who purchased Pro, Pro Plus, or Max through GitHub Mobile in-app purchase cannot buy additional AI credits from GitHub Mobile.

Grok Build states the general form: model access might vary depending on various factors such as geographical location and account limitations.

Sources: Anthropic, Plans and Pricing, individual plan cards, Japan-rendered state, read 21 August 2026 · AWS, Kiro Pricing, Common Questions, read 21 August 2026 · Anysphere, Cursor Models and Pricing, Start section, read 21 August 2026 · OpenAI, Codex Pricing, Feature availability section, read 21 August 2026 · GitHub, Plans and pricing, footnote 4, read 21 August 2026 · SpaceXAI Docs, Pricing, Billing and Availability section, last updated 3 July 2026

The product name is not the model name

Buyers evaluate a product name. Prices, context windows, and terms attach to a model name. On several of these products the two are separate objects, and describing them as one produces a comparison that the documents do not support.

The clearest case sits in the xAI documentation, and it is an object separation rather than a contradiction.

Kiro states that by default prompts are processed by Auto, described as an agent that uses a mix of different frontier models combined with specialized models for specific tasks, intent detection, and caching. It states that a task consuming X credits in Auto costs 1.3X credits through Sonnet 4.6. The unit of billing moves when the model choice moves, and the default is not a single model.

Antigravity’s individual plan card lists the available agent models as Gemini 3.5 Flash, Gemini 3.1 Pro, Gemini 3 Flash, Claude Sonnet and Opus 4.6, and gpt-oss-120b. Clause 8 of the additional terms states that a user selecting a third party or open source model as their main agent model will be subject to the terms of that model, and names Anthropic’s commercial terms specifically. The clause is additive on its face. It states that another company’s terms also apply, and nothing on the page states that Google’s own terms stop applying.

GitHub Copilot lists its available models on the plans page, and the list displayed under Free, Pro, Pro Plus, and Max is identical in all four columns. DERIVED: 32 entries counted in each of the four lists on the opened page. What differs between the four tiers is the credit balance and the entitlements attached to it, not the catalogue. Whether other GitHub surfaces publish the same catalogue was not checked for this briefing.

Cursor divides its models into two pools with different billing. The pool labelled Cursor Models contains Grok 4.6, Grok 4.5, and Composer 2.5, and carries significantly more included usage. Everything else is charged at the model’s API price from the Other Models pool. Section 11 records a corporate relationship relevant to reading that first pool.

Sources: SpaceXAI Docs, Grok Build Getting Started, read 21 August 2026 · SpaceXAI Docs, Models, last updated 18 August 2026 · SpaceXAI Docs, Pricing, last updated 3 July 2026 · AWS, Kiro Pricing, Common Questions, read 21 August 2026 · Google, Antigravity Pricing, read 21 August 2026 · Google, Antigravity Additional Terms of Service, clause 8, read 21 August 2026 · GitHub, Plans and pricing, Models section, read 21 August 2026 · Anysphere, Cursor Models and Pricing, Usage pools section, read 21 August 2026

A subscription is not a bucket of tokens

Two of the ten pages state that the subscription is valid only through the client software the vendor intends, and both name the same third-party harness as their example.

Google’s Antigravity additional terms, clause 6, state that using third party software, tools, or services to access the service is a breach of the agreement, give the example of using OpenClaw with Antigravity OAuth, and state that such actions may be grounds for suspension or termination of the account.

Kiro’s pricing FAQ lists the tools the subscription may be used with: the Kiro IDE, the Kiro CLI, Kiro on the web, Kiro Crew, ACP compatible integrated development environments, and automation in software development such as reviews during continuous integration and delivery. It then states that use through third-party automation harnesses such as OpenClaw, routing requests outside Kiro’s native interfaces, is not permitted.

DERIVED: pages stating a client restriction, 2 of 10, which is 2 ÷ 10 = 20.0 percent. A third page restricts a purchase channel rather than a client, and is counted separately in section 02.

Read alongside section 09, the consequence is narrow. Both vendors publish an open model list and a documented allowance, and separately define which program may spend it. The allowance is attached to a route.

Sources: Google, Antigravity Additional Terms of Service, clause 6, read 21 August 2026 · AWS, Kiro Pricing, Common Questions, read 21 August 2026

Who owns whom in this comparison

On 14 August 2026, Cursor published a post on its own blog stating that Cursor is now part of SpaceX. The post states that Cursor has been formally acquired by SpaceX, and that this completes an acquisition process which began with the announcement in April of a partnership with SpaceXAI to accelerate model training work. It states that the company will gain access to the largest GPU fleet in the world and the computing resources to build more capable models at lower operating cost, names Grok 4.6 released that week as an early example of what the two can achieve together, and links to the xAI announcement of that model.

Alongside that, the pages read here record the following. Cursor’s first-party model pool contains Grok 4.6, Grok 4.5, and Composer 2.5, and is the pool that carries significantly more included usage. The documentation site behind Grok Build titles itself SpaceXAI Docs, and its getting started page states that grok-4.6 powers the product. Cursor’s site footer names Anysphere, Inc. as the copyright holder.

What this briefing has not established, and does not assert, is the legal relationship between SpaceX, SpaceXAI, xAI, and Anysphere, Inc. The Cursor post uses two of those names in a single paragraph without defining the relationship between them, and no corporate filing was read. A reader who combines the four names into one corporate structure is supplying a link that these documents do not contain.

Because that link is not established, this briefing keeps the two products separate everywhere it counts. Cursor and Grok Build are counted as two products in section 02 and in every derived figure on this page. Merging them would require an official source defining control, and merging them without one would be the same error as the one described in section 09: supplying a relationship the documents do not state.

Two consequences remain worth naming. The first-party pool that Cursor prices most favorably is built on a model line that also appears in this comparison as a separate product, which is why the two Cursor pools are described separately in section 05 rather than merged into one figure. And this briefing carries no affiliate relationship, sponsorship, or commercial arrangement with any of the ten vendors, so no placement has been withheld or granted. There is no ranking on this page at all.

Sources: Anysphere, Cursor Blog, 14 August 2026 · Anysphere, Cursor Blog index, read 21 August 2026 · Anysphere, Cursor Models and Pricing, Usage pools section, read 21 August 2026 · SpaceXAI Docs, Grok Build Getting Started, read 21 August 2026

Where the published record stops

A document-first briefing is only as good as its account of its own limits. Each item below is limited to the page named and the date it was read, and none of them is a claim that the information is unpublished elsewhere.

Read but not fully extracted. GitHub Copilot’s FAQ entries on intellectual property and on plan differences were not expanded, which is why section 03 preserves a contradiction rather than resolving it. OpenAI’s Codex feature availability table carries rows for training and for retention and residency controls whose per-plan cell values could not be read reliably from the view recorded here, so no per-plan value is stated anywhere in this briefing. Antigravity’s asterisks could not be mapped to specific cells. Cursor’s third-party model price table ends with a control reading show more models, so the visible list is partial. Augment Code’s plan cards and its comparison table present the security items differently, and the visual state of the comparison marks was not confirmed.

Not read at all. The Cursor privacy and data governance page holding the eligibility rules for the ten percent residency uplift. Any Devin page beyond the pricing page, which means no training, retention, or security source for that product appears here. Any Command Code page beyond the pricing page, including the detailed pricing and limits documentation. Any Claude page beyond the general pricing page, including Claude Code product documentation. Any Kiro page beyond the pricing page. Any second GitHub surface publishing a model catalogue.

Stated but unquantified on the page. Command Code prints plus processing fee against every subscription price without stating the amount: NOT FOUND IN [commandcode.ai/pricing]. Google’s Antigravity pricing page displays no price for the Google AI Pro plan or the Google AI Ultra plan. Devin states that each paid plan carries an allowance refreshing daily and weekly without printing its size, and the five FAQ entries displayed on that page cover usage, reaching the limit, extending the limit, the Max plan, and the free plan, with none addressing model training.

Held behind an agreement. xAI’s security documentation states SOC 2 Type 2 status and that customers with a signed non-disclosure agreement can refer to its trust center for current certification and data governance information. Both that statement and Augment Code’s ISO 42001 and SOC 2 references are recorded here as vendor statements. No certificate, audit report, scope statement, or expiry date was read.

Naming and routing. Devin’s page uses Teams on the plan card and Team in the comparison table for what appears to be the same plan; both forms are preserved. The Codex pricing URL used here resolves to a page presented under ChatGPT Learn; the original address is retained as provenance.

Reading state. GitHub Copilot’s plans page was read in its Japanese locale, where the agent mode row displays 50 times per month for the Free plan. Whether another locale of the same page displays the same value was not checked, and no cross-locale claim is made. Anthropic’s pricing page was read in its Japan-rendered state, as labelled in section 08.

Sources: GitHub, Plans and pricing, Japanese locale, read 21 August 2026 · OpenAI, Codex Pricing, read 21 August 2026 · Google, Antigravity Pricing, read 21 August 2026 · Anysphere, Cursor Models and Pricing, read 21 August 2026 · Augment Code, Pricing, read 21 August 2026 · Command Code, Pricing, read 21 August 2026 · Cognition, Devin Plans and Pricing, read 21 August 2026 · SpaceXAI Docs, FAQ Security, last updated 27 July 2026 · Anthropic, Plans and Pricing, read 21 August 2026

Ten tools: the document map

Every claim on this page traces to one of the following, so that a reader can check the current state directly rather than take a dated reading on trust. The page count per product is shown because it bounds what section 02 could find.

DERIVED. Total pages: 1 + 1 + 1 + 4 + 1 + 2 + 1 + 4 + 2 + 1 = 18. Cursor and Grok Build together account for 4 + 4 = 8 pages, which is 8 ÷ 18 = 44.4 percent of the source inventory. Products represented by a single page: 6 of 10, which is 6 ÷ 10 = 60.0 percent. The distribution is uneven, and section 02 should be read with that in mind.

Sources: all eighteen pages above were read on 21 August 2026. Where a page displays its own last-updated date, that date appears in the section source line that uses it.

FAQ

Methodology

Tier. Tier C. Future Stack Reviews uses Tier C for briefings built from published documents with no hands-on testing. Tier B adds structured hands-on work, and Tier A adds long-term operational use. Tier measures how deep the evidence goes, not how good a product is.

What was done. Eighteen official vendor pages, listed in section 13, were opened and read on 21 August 2026. Where a page displays its own last-updated date, that date appears in the source line of the section that uses it. Where it displays none, the source line records the date it was read.

What was not done. No account was created, no software installed, no product run, no benchmark executed. No vendor was contacted before publication. No figure here describes speed, output quality, latency, or reliability.

The counting rule used in section 02. A cell is marked stated when a page opened for this briefing displays a value, a term, or an explicit restriction answering that question for that product. It is marked partial when the page states a relative or adjacent value rather than the value asked for, such as a sharing rule instead of a concurrency limit. It is marked not extracted when a relevant element was present but its value could not be read reliably from the view recorded. It is marked not stated when no such statement was found on the pages opened. Anyone can rerun this rule against the same eighteen pages and disagree with a cell. That is the point of publishing it.

Read values and calculated values. Any figure taken from a page is stated as what the page displays. Any figure produced by arithmetic, including every count, share, ratio, and multiplier, is labelled DERIVED with the arithmetic in the same sentence.

Absence. Where this briefing states that something does not appear, the statement is limited to the page named and the date read. It is not a claim that the information is unpublished. Section 12 collects these cases.

Source distribution. The page count per product is published in section 13 and is uneven. DERIVED: two products account for 8 ÷ 18 = 44.4 percent of the source inventory, and 6 ÷ 10 = 60.0 percent of products are represented by a single page. Section 02 is bounded by that distribution, and no cross-vendor frequency claim on this page reaches beyond it.

Product selection. Ten products were selected editorially from an internal working register of thirty-eight candidates. Products considered and deliberately excluded include Meta Muse Code, Mistral Vibe, Tabnine, GitLab Duo, JetBrains AI and Junie, Amp, Factory, Cline, Qoder, TraeCode, Kimi Code, the Z.ai GLM Coding Plan, Replit Agent, Freebuff, and Sourcegraph. Exclusion is an editorial decision about scope. It is not a statement that a product does not exist or does not qualify.

Relationship disclosure. Section 11 records a transaction announced by one of the ten concerning another. The two products are counted separately throughout, because the legal relationship was not established from any source read here.

Commercial position. Future Stack Reviews took no payment, no affiliate commission, no sponsorship, and no product access from any vendor named here. There are no affiliate links and no purchase buttons in this briefing.

Volatility. Prices, plan names, credit rates, model availability, and terms in this category change frequently, and several pages read here display last-updated dates in July and August 2026. Treat every figure as a reading of a specific page on a specific date, and open the linked page before committing a budget.

Corrections. Errors are corrected on the page with a dated note. Section 18 has the address.

Verdict

There is no winner on this page. The question a buyer faces is not which of these ten is best, but which tier of which one carries the rights the review process demands, and that answer moves with jurisdiction, team size, and who is asking.

What the eighteen pages support is narrower and more useful. Six ordinary buyer questions were put to ten products, and the published pages answered fifteen of the sixty. DERIVED: 15 ÷ 60 = 25.0 percent, rising to 22 ÷ 60 = 36.7 percent if partial answers are counted. No product answered more than three of the six, which is 3 ÷ 6 = 50.0 percent. The question about intellectual property indemnity was answered by one page out of ten, and that answer sits alongside other surfaces on the same page that this briefing could not read.

The seat price is not the least useful number on these pages. Nine of the ten pages read here display a headline price plainly, and Antigravity displays one for its individual route while leaving two plan cards without a price. DERIVED: pages displaying at least one headline price, 10 of 10; plan cards displaying no price, 2. Underneath it sit the objects that decide a procurement review, and on the pages read here those objects are more often absent, relative, footnoted, or split across surfaces than plainly stated. DERIVED: cells not plainly stated, 45 of 60, which is (7 + 3 + 35) ÷ 60 = 75.0 percent.

The practical instruction is short. Before signing in this category, open the plan comparison table rather than the plan cards, expand every footnote and every collapsed FAQ entry, and write down which of the six questions your page actually answers. If the count comes out near fifteen of sixty, that is not a failure of your reading.

Related briefings

Contact and corrections