Official offers, token-saving tips, and workflow guides.
Everything here was read on OpenAI’s, Anthropic’s, Google’s or Z.ai’s own page and written up with a link back to it. Nothing comes from a repost, a screenshot or a rumour.
API credits and pricing are not the same thing as a subscription allowance. Each entry names what it applies to, so read that before you plan around it.
Offers & programs
7
Z.ai·GLM Coding Plan · GLM-5.3-flashCheck availability
GLM-5.3-flash: bonus night-time usage extended through October 7
Paid GLM Coding Plan users get zero quota consumption in ZCode or double the usual quota in other supported agents during the nightly campaign window. Existing quota limits still apply.
ValiditySeptember 3 – October 7, 2026, with the nightly 23:00–09:00 window in Singapore time (UTC+8). The notice gives only calendar dates, so the exact closing instant is not stated; the end date was extended from September 20 to October 7.
The daily window runs from 23:00 until 09:00 the next morning, Singapore time (UTC+8), including weekends and public holidays. Benefits activate automatically.
ZCode usage requires version 3.10 or later. Other supported agents receive twice the allowance under the plan's normal quota rules.
An already exhausted five-hour or weekly quota can block participation until it resets. GLM-5.3 uses normal quota rules; this campaign is only for GLM-5.3-flash.
The end date was extended from September 20 to October 7, 2026.
Applies to
GLM Coding Plan · GLM-5.3-flash
Who it’s for
All paid GLM Coding Plan users selecting GLM-5.3-flash.
Z.ai·GLM Coding Plan · First subscriptionReferral offer
10% off your first GLM Coding Plan subscription
Register or sign in through our invitation link for a 10% discount on your first eligible GLM Coding Plan order. New accounts and existing users with no paid subscription history can qualify.
ValidityFirst qualifying subscription only; no campaign end date is stated.
Referral link — Reset may earn Z.ai credits if you subscribe. Eligibility and the final discount are confirmed by Z.ai at checkout.
What the source says
Open the invitation link and log in to bind the referral. Z.ai’s rules give a 72-hour window for a qualifying first payment to count toward the inviter’s reward.
The discount is available once per user. Renewals, upgrades, downgrades and similar first-order promotions cannot receive or stack this benefit.
The final payment must remain at least $0.50; Z.ai may adjust the discount to meet that minimum.
This referral benefit is separate from the GLM-5.3-flash usage campaign. Confirm eligibility and the amount payable at checkout.
Applies to
GLM Coding Plan · First subscription
Who it’s for
Z.ai users making their first paid subscription purchase.
OpenAI·Codex · ChatGPT Pro · APIApplication program
Six months of ChatGPT Pro for eligible open-source maintainers
Apply to the Codex for Open Source program for a complimentary subscription and support for maintainer workflows.
ValiditySix-month subscription benefit. No application deadline stated on the reviewed page.
The application lists several eligibility routes, including package adoption, project roles, and contribution history. Check the route that fits your work.
Contributors whose projects do not meet the listed thresholds can explain their ecosystem impact when applying.
At the end, an existing paid plan resumes at its previous rate unless cancelled; accounts without one return to the free plan.
Applies to
Claude Max · Claude Code
Who it’s for
Maintainers, core contributors, community builders, and critical-infrastructure projects.
Anthropic·Claude APIApplication program
Apply for Claude startup credits and priority rate limits
Claude for Startups offers a community program, with API credit benefits for qualifying funded companies.
ValidityEligibility and awards are subject to review. No application deadline stated.
For credit benefits, the reviewed FAQ requires institutional equity funding, a company founded within four years, and no previous Anthropic startup credits.
Prepare a Claude Console account, company email, website, and a description of the product. The page does not promise a fixed credit amount.
Credits apply to Anthropic’s first-party API through the Console, not a Claude app subscription, Amazon Bedrock, or Google Vertex AI.
Accepted startups are upgraded to Anthropic's highest API rate limits while on credits; standard pricing resumes automatically afterward.
Applies to
Claude API
Who it’s for
Early-stage companies building with Claude; credit eligibility is narrower than program membership.
Google·Google Cloud · GeminiApplication program
Up to $350,000 in Cloud credits for eligible AI startups
Google's AI startup program supports qualifying venture-funded companies over two years, subject to approval.
ValidityTwo-year program benefit. Funding-stage and application conditions apply. No application deadline is stated.
Inspect usage with /usage and context composition with /context to understand what the session is carrying.
Start a fresh session for unrelated work with /clear; use focused /compact instructions when the current task needs a shorter history.
Match model and reasoning settings to the task. A displayed API cost estimate is not the same as a subscription bill or remaining plan allowance.
Applies to
Claude Code
Who it’s for
Claude Code users managing long sessions or switching between unrelated tasks.
OpenAI·OpenAI APIOfficial guide
Reuse stable prompt prefixes with OpenAI prompt caching
Learn how matching prompt prefixes let OpenAI API requests reuse cached KV states, cutting input cost by up to 95% on reused tokens and reducing latency. Covers minimum cacheable lengths, implicit versus explicit breakpoints, cache retention, and cost measurement.
Prompt caching is enabled by default for supported models; reusing a cached prefix costs the reduced cached-input rate, discounted up to 95%, while a different model, changed tools, schemas, reasoning effort, or compaction can break prefix matching.
On GPT-5.6 and later, the minimum cacheable prefix is 1,024 tokens; explicit breakpoints via prompt_cache_breakpoint and prompt_cache_options.mode are supported, while earlier models support implicit caching only.
GPT-5.6+ cache writes cost 1.25x the standard input rate and cache reads cost 0.1x (0.05x on GPT-6.1 Sol); earlier models have model-dependent cached-input rates with no additional cache-write charge.
Control lifetime with prompt_cache_options.ttl (30m default) on GPT-5.6+, or prompt_cache_retention (in_memory or 24h) on earlier models; cached tokens still count toward per-minute rate limits.
Use a stable prompt_cache_key on models before GPT-5.6 to optimize routing, or optionally on GPT-5.6+ to keep separate cache accounting per customer or workspace; caches are never shared across organizations or regions.
Monitor cache hit rates via usage.input_tokens_details.cached_tokens, the Prompt Caching Dashboard, and by computing weighted input cost with cache-read and cache-write multipliers.
Applies to
OpenAI API
Who it’s for
API developers making repeated requests with shared instructions, tools, or conversation history on the OpenAI Responses and Agents APIs.
Anthropic·Claude APIOfficial guide
Cache reusable context in the Claude API
Anthropic documents automatic caching and explicit cache_control breakpoints for repeated instructions, tool definitions, and conversation context, cutting processing time and costs.
Enable automatic caching with a single top-level cache_control field, or mark a stable prefix with explicit cache_control breakpoints for fine-grained control.
Place reusable tools, instructions, and examples before request-specific content; cache writes happen only at the breakpoint, so changing an earlier block changes the prefix hash.
Use up to 4 cache breakpoints; the lookback window checks 20 blocks per breakpoint, so add a second breakpoint if a growing conversation pushes past the last cache write.
5-minute cache writes cost 25% more than base input tokens, 1-hour writes 2x, and cache reads 0.1x base input price (with per-model exceptions).
Compare cache_creation_input_tokens and cache_read_input_tokens usage and meet the per-model minimum cacheable prompt length; the cache refreshes for free on each hit.
For latency-sensitive apps, pre-warm the cache with a max_tokens: 0 request instead of the older max_tokens: 1 workaround.
Applies to
Claude API
Who it’s for
Developers using the Claude API with repeated prompt prefixes.
Google·Gemini APIOfficial guide
Improve cache reuse in the Gemini API
Google explains implicit context caching and how repeated inputs can receive cached-token cost savings.
Put large shared content early in the prompt and send requests with similar prefixes close together.
Check the model's minimum cacheable input size and inspect cached-token usage rather than assuming every request hits a cache.
The Interactions API supports implicit caching; Google directs explicit cache-object workflows to the generateContent API.
Applies to
Gemini API
Who it’s for
Gemini API developers sending substantial shared context across requests.
Google·Google Gemini API Batch APIOfficial guide
Gemini Batch API: Process Large Volumes of Requests at 50% Cost
Google's Gemini Batch API processes large volumes of asynchronous requests at 50% of standard cost, with a target 24-hour turnaround (often much faster). Suited to non-urgent workloads like data pre-processing and evaluations; jobs expire if pending or running over 48 hours.
Official sourceBatch APIai.google.dev · Checked 4 October 2026
What the source says
Submit requests via inline GenerateContentRequest objects (total under 20MB) or a JSONL input file uploaded through the File API (up to 2GB per file).
Batch jobs also support the Embeddings model via batches.create_embeddings for higher-throughput embedding workloads.
Poll job state via the returned job name; states include PENDING, RUNNING, SUCCEEDED, FAILED, CANCELLED, and EXPIRED (jobs pending or running over 48 hours expire with no results).
Webhooks can replace polling: subscribe to batch.succeeded / batch.failed events so the Gemini API pushes completion notifications to your server.
Batch job results are stored and available for download by default for 6 weeks before permanent deletion.
Batch generation is supported for image models (e.g., Gemini Nano Banana workflows), trading higher rate limits for up-to-24-hour turnaround.
Applies to
Google Gemini API Batch API
Who it’s for
Developers using the Gemini API who need high-throughput, non-urgent generation or embedding workloads and want to halve inference costs.
Google·Gemini API URL context toolOfficial guide
Gemini API URL context tool: fetch page content and track token usage
The Gemini API URL context tool lets models retrieve content from URLs you provide to inform responses, returning url_citation annotations. Retrieved URL content counts as input tokens, reported separately in tool_use_input_tokens within the usage object, so pricing follows the model used.
Official sourceURL contextai.google.dev · Checked 4 October 2026
What the source says
Enable the tool with tools=[{"type":"url_context"}]; it first checks an internal index cache, then falls back to a live fetch if the page isn't indexed.
Responses include inline url_citation annotations linking response segments to source URLs, plus a url_context_result step for debugging retrieval status.
Retrieved URL content is counted as input tokens: usage reports tool_use_input_tokens (e.g., 10309) alongside input_tokens; pricing depends on the model.
Limits: up to 20 URLs per request, max 34MB per URL, URLs must be publicly accessible (no localhost, private networks, or tunnels like ngrok).
Supported content: HTML, JSON, plain text, XML, CSS, JavaScript, CSV, RTF, common image types, and PDFs; paywalled content, YouTube videos, Google Workspace files, and video/audio are not supported.
Best practices: provide direct, complete URLs (with protocol) to publicly accessible pages; the model only retrieves from URLs you supply, not nested links.
Applies to
Gemini API URL context tool
Who it’s for
Developers building with the Gemini API who want models to read specific web pages, PDFs, or docs and need to account for the token cost of retrieved content.
Google·Gemini APIOfficial guide
Understanding Gemini API rate limits and usage tiers
Google's official guide explains how Gemini API rate limits work across RPM, TPM, and RPD dimensions, how spend-based limits apply per tier, and how projects automatically upgrade tiers as cumulative Google Cloud spending grows. It also covers Batch API limits, priority inference limits, and how to request paid tier rate limit increases.
Official sourceRate limitsai.google.dev · Checked 4 October 2026
What the source says
Rate limits are measured across requests per minute (RPM), tokens per minute (TPM), and requests per day (RPD), applied per project rather than per API key, with RPD quotas resetting at midnight Pacific time.
Spend-based rate limits are evaluated on a rolling 10-minute window: Free tier has none, Tier 1 is $10, Tier 2 is $50, and Tier 3 is $200 per 10 minutes; hitting them returns a 429 RESOURCE_EXHAUSTED error.
Tier qualifications: Tier 1 requires a linked billing account, Tier 2 requires $100 paid plus 3 days, and Tier 3 requires $1,000 paid plus 30 days from first successful payment, based on cumulative Google Cloud spending.
Free-to-Tier-1 upgrades typically take effect instantly and later upgrades within 10 minutes once criteria are met; billing is set up in AI Studio.
Batch API has its own limits including 100 concurrent requests, a 2GB input file size limit, a 20GB file storage limit, and per-model enqueued token caps that scale with tier.
Priority inference consumption holds separate default rate limits at 0.3x the standard rate limit for each model and tier, while still counting toward overall interactive traffic limits.
Applies to
Gemini API
Who it’s for
Developers building on the Gemini API who need to plan request budgets, manage tier upgrades, or resolve 429 rate limit errors.
Anthropic·Claude (Pro, Max 5x, Max 20x)Official guide
How to enable and manage usage credits for paid Claude plans
Anthropic lets subscribers on Pro, Max 5x, and Max 20x plans enable usage credits so they can keep working after hitting their five-hour plan limits, billing overage at standard API rates. Enable in Settings > Usage, add a payment method, set monthly spend caps or auto-reload, and prepay funds; a $2000 daily redemption limit applies. Mobile-app subscribers must enable credits on the web app.
Enable via Settings > Usage: click "Enable", set up a payment method, configure spending preferences, and use "Adjust limit" to set a monthly spend cap or "Set to unlimited".
Prepay funds with "Add funds" and optionally enable auto-reload when the balance falls below a threshold; there is a daily redemption limit of $2000.
Overage is billed at standard API rates and charged separately from the subscription; included plan limits still reset every five hours.
Usage credits apply to Claude conversations, Claude Code terminal usage, Research mode beyond plan limits, and all tokens processed including project file content.
Cost tips: use Haiku or recent Sonnet models, plan sessions around five-hour windows, start fresh chats, store documents in project knowledge, and set conservative spend caps.
Credits can be disabled anytime in Settings > Usage; in most cases they do not expire, but in certain jurisdictions such as Japan they expire six months after purchase starting September 10, 2026.
Applies to
Claude (Pro, Max 5x, Max 20x)
Who it’s for
Individual subscribers to Claude Pro, Max 5x, or Max 20x plans who want uninterrupted usage beyond their plan limits, including mobile-app subscribers and Claude Code users.
Workflow guides
6
OpenAI·CodexOfficial guide
Give Codex a clear goal, context, and definition of done
Describe the desired change, relevant files, constraints, and how completion should be checked.
Use planning for ambiguous or complex work; keep durable repository conventions and test commands in a concise AGENTS.md.
Turn established, repeatable workflows into skills. Evaluate changes against the task’s acceptance checks.
Applies to
Codex
Who it’s for
Developers using Codex for repository changes and repeatable coding tasks.
Anthropic·Claude CodeOfficial guide
Explore, plan, and verify changes with Claude Code
Anthropic’s practical guide shows how focused context, explicit verification checks, and staged explore-plan-implement workflows improve agent-assisted coding with Claude Code.
Give Claude a runnable check (tests, build, linter, or screenshot comparison) so it can verify and iterate on its own work instead of relying on your review.
For uncertain or multi-file tasks, explore in plan mode, write a detailed plan, implement against it, then commit and open a PR; skip planning for one-sentence fixes.
Point to concrete files, existing patterns, constraints, and tests rather than relying on broad requests, and provide rich context with @-file references, pasted images, or piped data.
Keep CLAUDE.md concise and reusable; move rarely used domain knowledge into skills loaded on demand to avoid bloating every conversation.
Manage context aggressively: run /clear between unrelated tasks, use subagents to keep research out of the main context, and rewind to checkpoints when approaches fail.
Scale with non-interactive claude -p runs, parallel sessions or worktrees, and adversarial review subagents that check diffs in fresh context before work is treated as done.
Applies to
Claude Code
Who it’s for
Claude Code users tackling unfamiliar code, multi-file changes, or unattended and parallel agent runs.
Google·Gemini API · Google AI StudioOfficial guide
Structure Gemini prompts around clear tasks and useful examples
Google's prompt-design guide covers clear instructions, few-shot examples, context, output constraints, parameter tuning, and Gemini 3-specific practices like consistent structure and long-context ordering.
Give clear, specific instructions and constraints; use few-shot examples to control format, phrasing, and scoping, keeping example formatting consistent.
For Gemini 3, be direct, use consistent delimiters such as XML-style tags or Markdown headings, and place critical instructions in the system instruction or prompt start.
For long contexts, supply all context first, put questions at the end, and bridge with a transition phrase like 'Based on the information above...'.
Break complex work into simpler components: separate prompts per instruction, chain sequential prompts, or aggregate parallel task outputs.
Keep temperature, top_p, and top_k at default values for Gemini 3.x models; changing them can cause looping or degraded performance.
Enable grounding with Google Search for obscure or recent facts and code execution for arithmetic, counting, or calculations.
Applies to
Gemini API · Google AI Studio
Who it’s for
Developers designing Gemini prompts and multi-step or agentic application workflows.
OpenAI·OpenAI APIOfficial guide
Move non-urgent API work to Batch
OpenAI’s Batch API offers a 50% cost discount relative to synchronous requests for supported asynchronous workloads.
Official sourceOpenAI Batch APIdevelopers.openai.com · Checked 30 September 2026
What the source says
Prepare requests in a JSONL file, submit a batch, then retrieve the completed results using each request’s custom ID.
Supported endpoints include responses, chat completions, completions, embeddings, moderations, and images generations/edits; each input file can only include requests to a single model.
The documented processing window is 24 hours. Check endpoint, model, and batch limits (up to 50,000 requests, 200 MB per file, up to 2,000 batches per hour) before moving a workflow.
Output files are deleted 30 days after a batch completes; batches that miss the window expire, unfinished requests are cancelled, and tokens from completed requests are still charged.
Batch changes delivery time and API pricing; it does not reduce an interactive Codex task’s subscription usage.
Applies to
OpenAI API
Who it’s for
API developers running offline evaluations, classification, or other jobs that can wait.
Anthropic·Claude CodeOfficial guide
Claude Code prompt caching: avoid cache invalidation and cut token costs
Anthropic's official guide explains how Claude Code uses server-side prompt caching to bill re-reads at the cached token rate while only new content is fully processed. It lists cache-invalidating actions (model switches, effort changes, fast mode, MCP server changes, plugin toggles, tool deny rules, compaction, upgrades), cache-safe actions, TTL defaults and overrides, and how to monitor cache hit rates.
Cache invalidating actions include switching models, changing effort level, turning on fast mode, connecting or disconnecting an MCP server, enabling or disabling a plugin, denying an entire tool, compacting the conversation, accumulating many images, and upgrading Claude Code.
Each model (and on most models each effort level) has its own cache; on Opus 5.5, Sonnet 5.5, and Fable 5.1 with an API key or a Claude subscription, changing effort keeps the cache intact.
Default TTLs: on a Claude subscription within plan usage the main conversation gets a one-hour TTL; with usage credits, API keys, or cloud providers it is five minutes. Set promptCacheTtl or CLAUDE_CODE_PROMPT_CACHE_TTL (v2.1.242+) to choose 5m or 1h yourself.
Cache-safe actions include editing repository files, editing CLAUDE.md mid-session, changing permission mode, changing output style, invoking skills and commands, running /recap, rewinding, and spawning subagents; skills append instructions as messages, keeping the cached prefix intact.
Run /compact at natural breaks between tasks rather than mid-task, and prefer /rewind when abandoning a path, since rewind truncates to an already-cached prefix.
Monitor cache performance via cache_read_input_tokens and cache_creation_input_tokens, or run /usage for a per-session Prompt cache (main) line showing hit ratio and likely miss causes (v2.1.251+, cause text v2.1.260+).
Applies to
Claude Code
Who it’s for
Claude Code users on API keys, Claude subscriptions, or cloud providers (Amazon Bedrock, Google Cloud Agent Platform, Microsoft Foundry) who want faster, cheaper sessions
Anthropic·Anthropic WorkspacesOfficial guide
Organize Claude API Usage with Workspaces
Anthropic's Claude Console lets organizations organize API usage into workspaces (up to 100 by default) to separate projects, environments, or teams while keeping centralized billing. Each workspace supports members, scoped API keys, workspace-specific rate and spend limits, and per-workspace usage and cost tracking via the Usage and Cost API.
Official sourceWorkspacesplatform.claude.com · Checked 4 October 2026
What the source says
Every organization has a non-renameable, non-deletable Default Workspace; it is omitted from List Workspaces and shows null workspace_id in keys, usage, and cost reports.
API keys can be scoped to a single workspace, or multi-workspace keys can target a workspace via the anthropic-workspace-id request header; responses return an anthropic-workspace-id header.
Workspace roles include Workspace User, Workspace Limited Developer, Workspace Developer, and Workspace Admin, plus an inherited Workspace Billing role; organization admins and billing members get automatic access across all workspaces.
Limits: workspace spend and rate limits can be set lower but not higher than organization limits, cannot be set on the Default Workspace, and default to organization limits if unset; the Claude Code workspace is the only one supporting per-user monthly spend limits.
Archiving a workspace deactivates it and archives its API keys within seconds, preserves reporting data, and cannot be undone; archiving the Claude Code workspace disables Claude Code Console-billing sign-in organization-wide.
Manage workspaces programmatically via the Admin API (Admin API key, org:admin OAuth token, or unscoped personal/service account key) under client.beta.organization.workspaces.
Applies to
Anthropic Workspaces
Who it’s for
Organization admins and developers managing Claude API projects, teams, environments, billing, and access control in the Claude Console or Admin API.
Get token offers and usage guides by email.
Join the list for the first issue. Up to two emails a month.
This is optional. Reset for macOS is free, and the download does not depend on it. The page stays here either way.