Docs changelog
The Docs section is versioned by date: every page carries an updated date, and this page lists what changed. The API's own versioning is on the Versioning reference page.
Requirements and headers gains Answers in the Certified pool: HIPS serves every request there at the pool's rates, the answer's x-hips-pool reads certified, and an answer there may carry no x-hips-gpu-* header. Models gains Request parameters: HIPS never strips a parameter, and where a model's sheet lists the parameters it takes, a request carrying another is answered 400 unsupported_parameter. Prompt caching: prompt_cache_key is accepted and never logged or stored.
Each published model has a sheet of its own at /docs/models/<id>, read from the catalog where it is served: its pools and the precision each serves it at, its limits, its rates and, where the model takes only some of them, the request parameters it takes (Models).
The provider Console: a GPU its machine reports paused while HIPS holds no pause of it reads Paused on its machine, with Resume where the machine kept a pause you ended in My GPUs, and where to resume a pause made on the machine itself.
With the Terms of Service 1.1 and the Common Clauses 1.1: Data handling opens with Your prompt data, "HIPS never trains on, retains or logs your prompt data.", what HIPS holds only where you ask and for how long, and what Prompt Data covers; Import and run your model states the commitment beside the weights commitment; Workspace settings and Deployments show it where the Console shows it; Billing quotes the Pricing table as it now reads, the plans named Free, Pro and Business with the commitment in every plan's row; and API keys names the Business plan for regions.
The provider Console: My GPUs lists machines and their GPUs by when they were added, the newest first, so none changes place while its readings change, and hides a machine HIPS has not heard from for more than 10 days until you show it.
Install and update hi: from 0.1.5 on, the macOS program is signed with HIPS's Developer ID and notarized by Apple, with the identifier com.hips.hi in every version, and the page shows how to check it.
Coding agents: from hi 0.1.6, Claude Code runs with tool search on, so your MCP servers' tools are sent only once Claude Code's search finds them; its /model list offers the session's model and the models whose window now holds 131,072 tokens, and the session is told the smallest window among them, so a switch to any of them fits; a Claude Code session needs 46,000 tokens, 52,000 before; and a model that no pool the session may use can serve now is refused, with the reason. Status and models: hi models no longer marks such a model fit for coding.
New guide Image generation and editing: the two image tasks, image_generation and image_edit, on the task API and on OpenAI's image routes; their options and bounds, input images and their metadata removed, the price per image at its size and what is billed, the answer, the content checks and the workspace's switch, the AI-generated mark, data handling, retries and refusals, the 120 seconds an answer's body must be read in, what a client that leaves is billed (the images made, those the output check would have withheld included when it left before the check finished) and that a retry with its idempotency key is then refused, idempotency with no answer kept, and what is not available yet. New reference page Images: POST /v1/images/generations in OpenAI's JSON shape and POST /v1/images/edits in its multipart shape, every field they take and refuse, output_compression as a jpeg output's quality among them, and the answer with its hips.withheld. Billing says how an image is priced, what a client that leaves before reading its images is billed, the output check's images included when it left before the check finished, and that a retry with the same idempotency key is then refused, and the 120 seconds an image answer's body must be read in.
Errors and the operating guidance's Errors gain content_refused, image_withheld and content_check_unavailable, and type_not_served names the edit route. Tasks lists the image tasks as priced, with their options, an image input's role and their answer; Usage the image unit; Requirements and headers idempotency-key on the image routes; the reference overview, Task guides and Models the image routes; and the compatibility matrix a row for each image route, unvalidated until its record.
Earnings and payouts gains Image work: a payout per image published at each size, beside its price, the one in force when the buyer's request was accepted; a withheld image paid, and the images made for a buyer who left; what earns nothing; a GPU's per-token minimum price not read for images; and image work with no published payout held under No published rate. The Provider setup overview names it among the page's topics.
The Prices API gains The price feeds per pool: GET /v1/feeds/certified/models, GET /v1/feeds/community/models and GET /v1/feeds/models on https://ai.hips.com, one machine-readable document per model and pool under the model id that runs it there, and the combined feed with the Community pool as the flex tier; and A pool in the model id: a model id without a suffix uses Auto, :auto pins Auto, :certified and :community pin one pool with no fallback, which chat completions, Requirements and headers and the API reference name too.
Messages serves tool search: a tool with defer_loading: true is not sent to the model until a tool_reference block in a tool_result names it, as Claude Code's MCP tool search sends them, and a context_management that only clears thinking is accepted. Both were a 400 before.
Errors and the errors guidance: a prompt held only in a size that the GPUs able to hold it keep failing to read within the time limit is context_length_exceeded, with no retry, since a retry would not succeed; context_size_not_measured stays for a size never tried or being measured. Both name the largest size that works now for the request, never more than the pool's context_window on Models. The same rows on Chat completions, Messages, Responses, Embeddings and Limits.
Earnings and payouts and Provider setup carry Irish VAT on the provider statements as it ships: the name Stripe verified, supplementary invoices and self-billed credit notes with their own objection, one VAT row per rate, and the months whose lines are on the next statement.
The Console no longer has a training switch, a training notice or a policy acknowledgment, and the Docs no longer describe them: Data handling drops the request-data policy by plan and Training on your request content, and Workspace settings, Usage, the Playground, Deployments, Batch jobs, Import and run your model, Authentication, Requirements and headers, Files and the Operating guidance overview no longer name a policy per plan.
A task, embeddings or rerank request's idempotency-key keeps no answer: a retry after the first completed answers 404 result_expired, its message naming the first request's req_ id, and nothing runs or is billed again (Idempotency, Tasks, Limits, the errors reference and Errors). A batch's input file is deleted once its batch ends, and a failed batch's findings quote nothing of a line (Batch jobs, Files). A lapsed or canceled paid plan, or an ended plan grant, moves the workspace to the free plan at once, with nothing paused and nothing to acknowledge; plan_lapsed is answered only by a workspace a lapse paused before then, for about a minute (Billing, Workspace settings, and the plan_lapsed row of the errors reference, of Errors and of Authentication).
Coding agents, Code with HIPS and each agent's page (Claude Code, Codex, OpenCode, HIPS Code, Kilo, Pi, Qwen Code, Crush, Cline, Prime Agent, Aider, Goose, Hermes Agent, Grok Build, DeepSeek Harness), for hi 0.1.3: auto mode, where a model reviews or approves an agent's actions, is not supported by hi yet. Each agent starts in its own nearest equivalent of Accept edits, edits in the working folder allowed and commands asking first, and a session started in a model-reviewed mode starts that way instead, with one line at launch; for Grok Build, whose auto mode a defaultMode of auto in your Claude Code settings would start, hi then leaves those settings out of the session. Each page names its agent's starting mode. The tested releases follow the signed list of 2026-11-14 (sequence 7), and the per-model results the record of 2026-11-14 (sequence 10): Claude Code passed with three models and Codex with two.
Coding agents, for hi 0.1.3: each model in Claude Code's /model list says the largest context window a request can get now on the GPUs serving it, and marks a model too small for a Claude Code session, which needs 52,000 tokens; one line at launch names such models. hi claude does not start Claude Code on such a model, naming its window and the models whose window holds a session, and passes over it when it chooses the model for you; asked for one directly with Claude Code's own --model, it starts it and says so in one line.
The Console: the first time you sign in, a short tour shows the two sides, Inference and Provider, and how to register a GPU; Take the tour in the account menu shows it again. Billing: the plan dialog opens each time the tour ends, at a first sign-in or after Take the tour, finished or skipped.
Semantic routing and Models: each router's hips.router in GET /v1/models names its type, the kind of model it chooses among, one of text, image, audio or video, and the tasks it routes; the four routers are text and route ["chat"].
Semantic routing is the new name of the Router (the company's request of 2026-11-13), in the sidebar, its own page's title and every link to it: Guardrails, Models, the Console, Models in the Console, Billing, Status and models and Data handling. The router names hips-router-chat, hips-router-agent, hips-router-code and hips-router-research, their suffixes and the page's address are unchanged.
New reference page Prices API: the public price reads on https://hips.com, with no key, for integrators who read HIPS's prices by machine. The catalog with every price row, the plans and the settings; one model's rates; and the price history, with the money in micro-dollars and a worked conversion, the units, the pools, scheduled prices, caching and the errors. The API reference lists it, and Billing points to it. The three routes name the pools certified and community, as the inference API does.
The Console says what each side of the workspace switcher is for, and that Earn with your GPU on the Inference side and Run AI models on the Provider side open the switcher with the other side marked.
Workspaces and members gains Your first workspace: HIPS makes it with your account, even for a GPU provider, and buying credits, plans, keys, the Playground and a tool's sign-in wait until you complete its details, which accept the Terms of Service. Your account says the sign-up step completes that workspace.
Errors, Handling errors and Chat completions gain 503 model_loading: every GPU that holds the model would have to load it first, and none could give a first token within the request's bound now, so the request is not sent to one and HIPS's placement brings the model up for the requests that follow; retry after Retry-After, nothing charged. The Router tries its next model on it, as on no_capacity.
Coding agents, for hi 0.1.2: hi tells each agent the context window its requests can use, so the agent compacts before a request outgrows the model; it offers and starts only models your workspace's plan and guardrails allow; and hi claude starts Claude Code in its Manual mode until HIPS has measured a model for auto mode's review, unless you name a mode.
Messages takes a system message in messages, as Anthropic's mid-conversation system messages: a string or text blocks after a user message, followed by an assistant message or ending the list, sent to the model as an instruction at its place. It was a 400 before.
Provider setup: the Settings window's parts start with the Overview and Achievements, both arriving with HIPS Provider 0.1.26, and with that version the app's menu opens the window on its Overview with Show my GPUs, on the settings with Settings, and your earnings in the Console with Show my earnings.
Provider setup gains The Settings window: its parts, and the two that arrive with HIPS Provider 0.1.26, the Overview with its ten layouts and the Achievements of your provider account. The Command reference gains hips-provider overview, for HIPS Provider 0.1.26: what the Overview of the Settings window shows, printed as one JSON object, counts, sizes and times only.
When your computer works: gpu_load and cpu_load read "Other programs are using the GPU heavily" and "Other programs are using the processor heavily", as HIPS Provider 0.1.25 says them, and in Always a pause made only of other programs' load shows Paused.
Prompt caching says what is true today: HIPS keeps no prompt cache yet, cached_tokens is 0 on every answer, and every input token is billed at the input rate. It also says what HIPS has decided: a workspace's prompt prefix kept only in the memory of the GPU engine that ran the request, in both pools, reused only by the same workspace, dropped after 30 minutes without use, first for speed and with no change of price. Batch jobs, Billing, the usage reference and the Messages reference say the same.
The Settings reference: update_rule says that Check for updates in the app installs the newest version at once, without waiting for automatic updates to reach this computer.
The vector store routes refuse a write while the workspace's stored files are marked for deletion with 400 storage_marked_for_deletion, the public API's status for a request the caller can change; the Console's own routes answer the same code as a 409. Billing, the errors reference, what to do about each error and Vector stores say so.
Models and the Router: the Models screen's list is in tabs by model type, Text, Image, Audio, Video, Decision and Specialist, and opens on Text, with no tab of every model; a router is listed under the tab of the models it chooses among, so the four routers, which choose among language models, are listed under Text alone. The Decision tab says that no decision model is published yet and what one is. The tab chosen stays in the address, and a search names the other tabs where it finds models.
Network & capacity, and its row in the Console, show always-on capacity alone: the per-model availability view is gone, since Models shows each model's availability in each pool, and the page offers no pool, pricing mode or fallback, since always-on capacity comes from the Certified pool alone.
New group Code with HIPS: HIPS Harness, your own coding agent started on HIPS models with one command and no API key; the installers, the first steps, the fifteen agents and their pages, which models pass each agent's tests, billing, and what is kept of a session.
Coding-tool setup lists every agent hi starts with its command and its page, and the supported-tools matrix gains a column for each tool's hi command.
Every agent hi starts has a page of its own, opening with the quickest way, hi, before its manual setup, with the releases hi starts on each system and the models that passed its tests: Claude Code, Codex, OpenCode, HIPS Code, Kilo, Pi, Qwen Code, Crush, Cline, Prime Agent, Aider, Goose, Hermes Agent, Grok Build, DeepSeek Harness; Other coding tools and editor integrations moves after them.
Billing gains Weekly pool of included credit and Changing plans: a paid plan's included credit comes as a pool each week, the monthly amount times 12 over 52, added while the period it starts in is paid, paying a negative balance first and expiring at the week's end with no rollover, with notices at 75% and 100%; an upgrade takes effect at once with the difference charged for the time left and the week's pool topped up for the rest of the week, and a downgrade waits for the period's end. A 402 insufficient_credit on a paid plan carries pool_refreshes_at, the next pool's start, on Billing, in what to do about each error and in the errors reference.
Install and update: the installer signs you in with hi login once hi is installed, when you run it at a terminal and this computer is not signed in yet; HI_NO_LOGIN=1 skips it, and it is skipped in CI and without a terminal.
Coding agents: whatever would send your session to the agent's vendor stays off whatever --keep-agent-telemetry says, and sharing a session is off for OpenCode, Kilo and HIPS Code; the flag keeps the agent's usage metrics.
Status and models and Coding agents: HIPS's router for coding agents, hips-router-code, is the one router hi models lists, first, and the model hi chooses after the one you last used, while it has models to choose; Claude Code's /model and OpenCode's model list offer it too.
Earnings and payouts says that earnings whose Irish VAT comes to less than half a cent are booked as a line of no VAT, so a month's remainder is kept by its booked lines until the month's next line carries the cent, with no document of its own, and that when a correction leaves an amendment's VAT at nothing, the correction answers an objection to it; that a month whose only rows are lines of no VAT has no statement, Statements saying so for the month; and that an adjustment of an Irish VAT line is stated as a change of VAT, never of the supply.
Provider setup lists Earnings and payouts with its supplementary invoices and self-billed credit notes.
Billing gains When storage cannot be paid: a month's storage that leaves the balance below zero marks the workspace's stored files for deletion 30 days on, with the notices at once, 7 days and 1 day before, nothing new added meanwhile, the credit that clears it at once, and what is deleted on the date. The errors reference and Vector stores list 409 storage_marked_for_deletion.
The provider Console gains Your devices at a glance: the gates' counts over one lane per GPU state, each lane's reasons with who acts on each and what ends it, a machine lit up in every lane, and the machine a GPU opens beside the lanes.
Billing gains Plans in the Console: Free, Pro and Business side by side on Free, monthly or yearly, with each plan's weekly pool of included credit, models and features read from the plans' terms; the line under the header on Free; the plan dialog at an owner's or admin's first visit to a workspace; a locked model's or feature's reason beside it, on Models and on the Router, Guardrails and Knowledge screens; and this week's included credit on Pro and Business, with its 75% and spent lines and the ways on.
Earnings and payouts says when Stripe has verified your name: once it lists nothing more it needs about your identity, whatever its deadline, and your account can receive transfers or payouts; until then no Irish VAT is added and Settings and the invoice say "Stripe is still verifying this account; VAT is added from the first statement after it finishes". It also says that the invoices and the supplementary invoices and credit notes state the VAT in one row per rate, with its base; that a month's sub-cent remainder waits across its invoices for the line that tips the rounding, whose row then says the rounding it carries; that after a change of either name Settings shows the number as being checked; and that an objection HIPS had not answered moves to the document that replaces the one objected to when HIPS corrects the invoice it amends.
Earnings and payouts says that VIES must also give the name Stripe verified on your Connect account, that only the part of VIES's name before a "T/A" counts, and that a change of either name stops the Irish VAT until VIES confirms it again; that the VAT of earnings goes with the invoice that holds them, at the rate in force on the last day of their month, with no sub-cent remainder lost; and that each supplementary invoice or credit note has its own Object and review, and is reissued naming the new invoice when the invoice it amends is corrected.
Earnings and payouts gains Supplementary invoices and credit notes: Irish VAT is booked by the month of the supply, an adjustment in the month of the earnings it adjusts and a reversed line's VAT taken back at the next settlement, and VAT of a month whose invoice was issued already comes as a supplementary invoice or a self-billed credit note naming that invoice, issued with the next statement and accepted or objected to with its invoice. It also says that VIES must give your payout profile's legal name for an Irish VAT number, how the names are compared, and that a changed legal name is checked again by itself.
Earnings and payouts gains Settlement Statements and self-billed invoices: a statement for each month by its Payable Date, what it shows, its PDF on GPU earnings, the self-billed invoice HIPS issues in your name with its number, the euro figures at the European Central Bank's reference rate, the reverse charge or Irish VAT, the thirty days to accept or object, silence as acceptance, and the review of an objection. The Provider setup overview names it among the page's topics.
Chat completions and Responses: parallel_tool_calls: true is a permission, as OpenAI means it, never a requirement. A model with tools that does not advertise parallel_tool_calls serves the request as if it said false, at most one tool call a turn, where it answered 400 unsupported_parameter before; false is accepted on every model with tools. The Router no longer narrows its candidates on it.
The Provider setup overview lists Data centers.
Billing says that a second payment taken for an auto reload already paid is refunded in full at once.
Billing says where a plan is canceled and resumed: on the Billing screen, under Plan; Stripe's billing page changes the card and shows the invoices only. Choose a plan gives the platform's own reason beside a plan it cannot sell by card.
Earnings and payouts gains Holds on your earnings: No published rate, for work that ran with no published rate, kept in full until HIPS prices it.
The provider Console gains It is being checked again: a GPU whose passed check rested on tests that no longer count is checked again by itself, and its page says why and since when.
The plans' gates: a model or feature the workspace's plan does not include answers 403 model_not_on_plan with feature and required_plan, on the errors reference with what a workspace keeps below a feature's plan and in what to do about each error, and in the lists of chat completions, tasks and deployments. Models marks a model or router the plan leaves out in hips.allowed_by with the plan and the lowest plan that includes it.
Earnings and payouts gains Earnings as Credits: a share of each month's payout, or what is payable now, taken as Credits in a workspace you own alone once the Additional Terms are published and accepted. The Provider setup overview names it among the page's topics.
When your computer works gains What Always lifts and What still pauses work in Always: from the next app release, Always no longer stops for the microphone or the camera in use, a full-screen app, a game, a video playing or being encoded, a remote session, Do Not Disturb or a missing session helper, while heat, the CPU throttling, a driver reset, battery and low-power mode, the daily cap, a pause and other programs' load on the GPU or the CPU still stop work; Auto is unchanged. A lifted sign stays on the computer, shown there as "Active, not pausing work in Always", and is never sent to HIPS. Every reason now says what it does in Always, and Always's description names what still pauses work, in the app's own words. The settings reference quotes the app's help for Always in availability_mode and says the same.
Errors holds both of the day before's changes at once: the 402 row's auto_reload, and the guardrails' rows for a request that requires a region.
Coding agents says what is kept of a session: HIPS keeps no prompt or answer, and the agent keeps its own history on your computer, as it does without HIPS. The notice before a workspace's first launch names the pool its requests run in, and no request-data policy.
Guardrails gains A request that requires a region: the checks a decision service judges are not run for such a request, a block or redact one answering 503 guardrail_unavailable and a flag one reported unavailable with reason region, the same once the workspace's own share of the day is spent (workspace_share) or it holds its part of what is left of the day (day_share), and the Console says so beside each such check. A redact check that could not run for any reason now refuses the request as block does (decided 2026-11-11 by the company). Data handling says such a request sends the decision service nothing, and the errors reference that its guardrail_unavailable is not worth a retry.
Auto reload is built: Billing says how the owner turns it on, its line, amount and monthly limit, what happens when a payment fails or a month reaches its limit, and a consumer's express request. A 402 insufficient_credit carries auto_reload (on, capped, failed, off): Billing and Errors.
The command line covers hi as released: install and update from hips.com/hi.sh, hips.com/hi.ps1 or Homebrew, the options of `hi login`, `hi status` and `hi models`, and the coding agents hi starts on HIPS, with their options and what happens at launch.
New group Console: your account, workspaces and members, API keys, workspace settings, usage and activity, models and Network & capacity, the Playground, and deployments and imports.
New group Billing: plans, credit, top-ups, budgets and alerts, how a request is priced, storage, receipts and disputes.
New group Provider setup: when your computer works, the command reference and the settings reference of HIPS Provider 0.1.22, the provider Console, earnings and payouts, and crash reports.
The Command line lists the commands hi has; sign-in is served, a tool stays signed in until it is revoked, and Account's list of authorized devices and apps is described. Batch jobs gains In the Console, and Data handling gains Training on your request content. The HIPS SDK overview names no package or method, and the eight per-SDK pages are unpublished until the SDKs exist. This changelog itself is dated today.
New groups Router and Guardrails. Router: the four router names, the suffix, how a model is chosen, x-hips-router and the hips object's router, the bill, the refusals and the Console's Router page. Guardrails: the three types and their settings, assignment and how several combine, what a client receives on each route, redaction, the settings the workspace already had, limits and data handling. Both are live in production, the Router's read and the guardrail checks a decision service judges included; each page says what a request gets when that service has no answer. Data handling gains Routers and guardrails, linking both.
The API reference says what production serves, read from the gateway's code: the overview lists every route with its state in production (served, not offered, not served, not built) and no longer says that no route answers; Responses is served and documented whole; Jobs are not built and answer 404 route_not_found, as the task guides now say (classification, image, video, audio, document processing), the image and audio guides adding that their OpenAI routes answer type_not_served; Uploads are not offered yet and add DELETE; Deployments says chat completions serve no deployment yet; Models lists the routers first and the members the route writes (tool_calls, pricing_modes, hugging_face_id, allowed), and no longer names available_on_plan, which it never wrote; Tasks lists the task types and their units; Usage bills embeddings and rerank per started second (H109) and says no cached token is billed yet; Requirements and headers adds every header since 2026-10-21 and the request headers of Messages; Errors lists every code the gateway answers, with Messages' error shape. Chat completions adds prompt_cache_key and the refused cache fields, Authentication what each key kind reaches, Limits the 16 stop entries, Idempotency the routes that take a key, the compatibility matrix rows for Messages, embeddings, rerank, files and batches. Batch jobs says batches are not offered yet, with the validation codes; Prompt caching says what is served now; Sign in is served; Import and run your model adds architecture_unsupported and hub_token_unavailable. Every page now carries the date it last changed.
New reference pages: Messages, Anthropic's Messages API with its headers, events and error shape, and Count tokens, not served yet; Embeddings and rerank, billed per started second; Files, Batches and Webhooks, built and not offered yet; Vector stores. A new guide, Knowledge, built and not switched on in production yet, its words new copy for approval. Operating errors adds the public status line, GET /status.
The Console's Overview, and its page Console, state the API root as it is in the first request's second step ("Reword as live").
The Quickstart states the API root as it is and is called Quickstart, not Quickstart preview, in the sidebar, on the Developer hub, on Migrate to HIPS and in the site's footer ("Reword as live").
The Developer hub says the pages document the HIPS API as it runs in production, and the Quickstart no longer calls itself a preview or its examples a future interface contract (decided 2026-11-11 by the company, "Reword as live"). Every page ends with a link to sign in to the Console in place of "Join early access", which stays in the site's header ("Replace it"). The Vector stores reference is approved with the Knowledge guide ("Publish with the guide").
No page carries a preview banner any more, and llms.txt and llms-full.txt no longer open with the prelaunch statement (decided 2026-11-11 by the company, "Remove the banner"): each page says in its own text what is and is not available. The task guides that name the job routes say near their top that those routes are not built (classification, image, video, audio, document processing). Knowledge's words are approved as written ("Publish it now").
Batch jobs lists budget_exceeded, a batch a budget guardrail cannot carry; Chat completions says a router name asks HIPS to choose the model; Usage says a routed request is billed as the model chosen, with no router fee.
A machine connected with an enrollment code stays connected when you use Revoke all, which ends your own tool sign-ins, and End all machines of a code signs it out: your account, sign-in and the provider Console.
Request-data training is off by default on the lower paid plan, as on the higher, and an owner or admin of a workspace on either plan may turn it on: Billing, workspace settings, Data handling and Import and run your model.
New guide: Data centers, for operators who run HIPS Provider on servers: installing on a server and in Docker, the data-center flag, pause and resume, handing GPUs back to your own clients, automating it and monitoring it.
The verification codes are not sent while identity verification is switched off, and account_restricted is the refusal sent instead: Authentication, Errors, operating errors, Sign in.
A prompt's size: context_window per pool on Models, context_size_not_measured and the count in context_length_exceeded on Errors and Limits, the words of a no_capacity answer; the exact model of the GPU that served an answer on Streaming (H132).
What each pool can do for a key now (availability), the coding agents a model works with and the list's defaults on Models; the served GPU in the hips object on Streaming; what a no_capacity answer says on Errors.
The pools are the Community pool and the Certified pool, community and certified on the wire, the former names still taken: Requirements and headers, Deployments, Jobs, Data handling, Import and run your model, Batch jobs, Prompt caching, the HIPS SDK and its language pages, the compatibility matrix.
Auto is the default pool: the requirement takes auto, and x-hips-pool names the pool that ran the work (Requirements and headers, Tasks).
An undecided value reads "not decided yet" on every page; the Usage reference bills a cached prompt token at the cached-input rate; the Models page and the task guides say that limits and the task list are set per model in the catalog.
The base URL https://ai.hips.com/v1 is in every sample (decided 2026-10-21). New guides: Batch jobs and Prompt caching, both previews of decisions taken the same day; a page per SDK for the eight languages, from Python to Ruby; the Command line group with sign-in by device authorization. The Authentication reference carries the decided sign-in lifetimes.
Customer models are imported from a model hub, never uploaded (decided 2026-10-21). Upload and run your model is now Import and run your model, at a new address, written as the import's ten sections: what a customer model is, hubs, access tokens, formats and refusals, revisions and versions, licenses, checks and profiling, pools, deploy and call, weights and request data. The Developer hub and the HIPS SDK carry the new import sentence; the Uploads reference is for task inputs only; Data handling adds the import path, the hub token's handling and that no customer code runs; the Deployments reference adds the tariff_pending state.
First preview of every page, written against the source specification's developer copy and the inference API contract: the Developer hub, the Quickstart preview, Migrate to HIPS, the HIPS SDK, Upload and run your model, Models and the six task guides, Coding-tool setup with the supported-tools matrix, HIPS Code as its first row, and one page per published row, the twelve API reference pages, the seven operating guidance pages. Every page is a preview that links to early access; every matrix row is unvalidated.
- Sheet
- 20 of 20
- Updated
- 2026-11-16