July 1 decision: test Sonnet 5 before Opus API prices

Claude Sonnet 5 is the new first test for routine Claude coding and agent work. Through August 31, 2026, its API price is $2 input / $10 output per 1M tokens, moving to $3/$15 afterward, versus $5/$25 for Opus 4.8. Anthropic positions Sonnet 5 as close to Opus 4.8 and says it can match Opus on some tasks at higher effort; those are vendor claims, so retain Opus for highest-accuracy review and arbitration until your own tasks show otherwise. Sonnet 5’s new tokenizer produces about 30% more tokens for the same text than Sonnet 4.6, which can reduce equivalent context and change per-request cost. Re-count real prompts instead of multiplying old token estimates by the new list price.

GLM-5.2 remains the best low-cost supported-tool coding plan to test. Its API anchor is $1.40/$4.40, while the $18+ GLM Coding Plan applies only through supported tools and shared subscription quota. The mechanics matter: GLM-5.2 and GLM-5-Turbo deduct 3× during peak hours and normally 2× off-peak; a temporary benefit reduces off-peak deduction to 1× through the end of September. Five-hour and weekly caps still apply. Use the Z.AI buyer guide for plan sizes, supported tools, and renewal rules rather than treating the subscription as a general API allowance.

AlternativeCurrent roleRead before paying
Kimi K3Latest Kimi flagship; 1M context; $3 cache-miss input / $15 outputOpen weights are pending until July 27; run CAR tests first
Kimi K2.7 CodeCheaper routine Kimi coding API; $0.95 cache-miss input / $4 output256K context; thinking is required; HighSpeed costs more
MiniMax M31M coding-agent and $20 Token Plan testVendor benchmarks; quota and PAYG are separate
Xiaomi MiMo V2.5Low-cost hosted API, Token Plan, and MIT open weightsHosted privacy and credit burn need local checks
Gemma 4Private local/offline laneHardware fit matters more than cloud token price
Qwen3.7-PlusAlibaba’s lower-cost balanced coding optionSingapore: $0.40/$1.60 up to 256K input; $1.20/$4.80 above
Qwen3.7-MaxAlibaba’s strongest reasoning/coding optionSingapore: $2.50/$7.50; not the budget Qwen lane

Alibaba prices above are explicitly for the Singapore international deployment and are per 1M tokens; other regions differ.

July API value shift

Low-cost supported-tool/API lane

GLM-5.2

$1.40 in / $4.40 out

1M context

Vendor specs + independent Artificial Analysis signal

Claude cost/performance test

Sonnet 5

$2 / $10 through Aug 31

$3 / $15 after · 1M context

Anthropic vendor evidence; independent normalized result pending

Premium accuracy and arbitration lane

Opus 4.8

$5 in / $25 out

1M context

Vendor specs + independent Artificial Analysis signal

USD per 1 million tokens. Prices, context limits, and evidence provenance are separate decision inputs; this is not a benchmark ranking.

Subscription seats are a separate lane

As checked July 1, 2026, both OpenAI and Anthropic publish $20 regular-use seats, $100 5× tiers, and $200 20× tiers. The multipliers are relative to each provider’s base paid seat—not API tokens—and individual models or coding tools can still have separate five-hour, weekly, or model-specific limits.

Provider$20 regular-use seat$100 tier$200 tierLimit caveat
OpenAIChatGPT Plus for lighter usePro 5× PlusPro 20× PlusCodex promotions and model-specific quotas can differ
AnthropicClaude Pro standard capacityMax 5× ProMax 20× ProSession, weekly, and model-specific limits still apply

Account-level UI evidence showing a timed reset or extra-usage option does not establish universal weekly limits or repeated bonus resets. Use the live account meter and provider help pages for the applicable seat.

Not paying yet? Start here: Free Frontier Stack

Claude and ChatGPT: The Squeeze

Anthropic still draws a hard line between first-party subscription usage and third-party API usage. Claude Code and claude.ai remain the subscription-native path; third-party harnesses should be planned around API keys and API pricing. Current Claude API anchors list $2/$10 introductory pricing for Sonnet 5 through August 31, $5/$25 for Opus 4.8, and $10/$50 for Fable 5.

Fable 5 belongs above Opus in the escalation ladder, not inside the value stack. Anthropic says the June 12 export controls were lifted on June 30 and restored Fable globally on native Claude surfaces from July 1. Included usage is temporary and plan-specific, cloud re-enablement is still rolling out, and guarded-domain behavior and 30-day retention remain deployment considerations. Treat Fable as an expensive tested escalation—not routine routing. The longer decision guide is here: /posts/claude-fable-5-mythos-5-cost-guardrails/

GPT-5.6 Sol, Terra, and Luna also have published preview prices, but OpenAI says access starts with a select trusted-partner cohort. Do not treat those prices as generally available value routes. Keep GPT-5.5 as the active OpenAI baseline until your account is eligible and the model-specific retention, guardrail, and fallback terms are verified. See the frontier access-gates analysis.

OpenAI’s current GPT-5.5 help page is more useful than stale “capacity tier” shorthand. It lists GPT-5.5 Instant as available to logged-in users, Plus/Go users at up to 160 GPT-5.5 messages per 3 hours, Plus and Business users at up to 3,000 manually selected GPT-5.5 Thinking messages per week, and Business/Pro as unlimited for GPT-5 models subject to abuse guardrails. ChatGPT Plus remains a $20/month web-app subscription; API usage is separate.

The practical pattern is clear: use flat-rate apps where the provider supports your workflow, but do not assume a subscription covers third-party coding tools or API-heavy automation.

Cost comparison: /compare/claude-vs-openai/pricing/

Free Frontier Access (Verified)

You do not need to pay immediately to access strong models. These options were rechecked on April 8, 2026:

AMP — Ad-Supported Opus 4.5 (No Credit Card)

Sourcegraph’s AMP offers $10/day in free credits via an ad-supported tier—non-intrusive text ads in exchange for sustainable access to the $200/month model. Credits replenish hourly (~$300/month equivalent).

Best for: Developers who want predictable, ongoing Opus 4.5 access without worrying about preview terms changing (Antigravity) or credits expiring (Kiro). The ad experience is minimal (text-only, never influence responses) and daily limits are sufficient for focused coding sessions.

The catch: CLI execute mode and SDK require paid credits (not available on free tier). Data sharing for training no longer required.

Tool deep dive: /tools/amp/

Source: ampcode.com/news/amp-free-frontier

Antigravity — Free Opus 4.5 (No Credit Card, No Ads)

Google’s AI-native IDE offers free Claude Opus 4.5 access during public preview—the $200/month model with no credit card required. Also includes Gemini 3 Pro, Gemini 3 Flash, and agent workflows.

Best for: Sustained free access to a historical Opus-class reasoning lane without ads. No credit card required.

The catch: Public preview means terms can change. Rate limits apply. Data handling differs from Anthropic’s direct service.

Tool deep dive: /tools/antigravity/

NVIDIA NIM — Free Kimi k2.5 API

FREE

NVIDIA’s model hosting platform offers free trial credits for Kimi k2.5 API access—no credit card required. This is still the clearest answer when your question is specifically “where do I get free Kimi via API?”

What you get:

  • Kimi k2.5 via API: 76.8% SWE-bench, 256K context, OpenAI-compatible format
  • Thinking + Instant modes: Full model capabilities
  • OpenClaw-compatible: Kimi is explicitly allowed for third-party tools
  • No credit card: Trial tier requires only NVIDIA account signup

How to claim:

  1. Visit build.nvidia.com and create free account
  2. Search “Kimi K2.5” and click Get API Key
  3. Use with base URL https://integrate.api.nvidia.com/v1

The catch: Trial credits are limited (exact quotas not published). Not for production high-volume use.

Best for: OpenClaw users needing a fallback model, developers wanting API access without IDE overhead, Anthropic/Google migrants seeking ban-safe alternatives.

→ Complete NVIDIA NIM setup guide →

Last verified: April 8, 2026

OpenCode — choose the provider and billing path

OpenCode is an open-source coding harness, not a permanent free-model bundle. Current Zen docs describe an optional hosted gateway with billing details, credits, and per-request/model pricing. Promotional free rows can change.

For GLM-5.2, keep three lanes separate:

  • OpenCode Zen: optional OpenCode-hosted gateway and Zen pricing.
  • Direct Z.AI PAYG: standard Z.AI API billing.
  • Z.AI Coding Plan: subscription quota through the officially supported OpenCode integration.

Compare Pi, ZCode, and OpenCode with the same model and task before assuming the cheapest token or largest context produces the cheapest successful patch.

OpenRouter — Qwen 3.6 Plus Preview

Before you move to paid upgrades, it is worth running your eval set against Qwen 3.6 Plus Preview through OpenRouter. The appeal is simple: free access and a 1M-context preview without committing to a subscription first. The limit is also simple: if privacy, stability, or repeatability matters, move to a paid or official path after the first test pass. See Free Frontier Stack for the broader free-tier context.

Google AI Studio — Free Gemini Lane

Google’s current docs no longer support a simple “5-15 requests per minute” buying rule for the Gemini app experience. Gemini Apps describe usage as plan limits plus compute pressure: more advanced models, higher thinking levels, media generation, Deep Research, and larger context consume more of your usage.

On the API side, pricing and availability are model-specific. Gemini 3.1 Flash-Lite Preview currently has free-tier token pricing and a paid Standard anchor of $0.25 input / $1.50 output per 1M tokens for text/image/video. Gemini 3.5 Flash also lists free-tier token pricing, with a higher paid Standard anchor. Preview models may have more restrictive rate limits.

Best for: High-context analysis, codebase review, document processing, and cheap API experiments where the exact Gemini model and tier match your workload.

Trade-off: Free tier data can be used to improve Google products. Enable billing when privacy terms, repeatability, or production use matter.

Complete free tool breakdown: Free Frontier Stack

Coding Tools: Where the Value Is

Cursor — Still the Standard

Credit-based model: 50 requests/month free, Pro ($20/mo) gives you a $20 credit pool. The generous trial era is over, but it’s still the most polished AI coding IDE.

Windsurf — Avoid for Now

The product’s future is unclear after the acquisition collapse. If you loved Windsurf, you’re in limbo. Don’t build new workflows around it until Cognition clarifies the roadmap.

OpenAI Codex

Available to paid ChatGPT users (Plus/Pro/Team). Agentic coding for long-horizon work, large refactors, multi-file changes. Install via npm i -g @openai/codex.

Kiro CLI — 500 Free Credits (Verified)

New accounts get exactly 500 bonus credits (30-day expiration) usable on free tier models.

How to claim:

  1. Install: npm install -g @kiro/cli
  2. Create account during setup (credit card required)
  3. Check balance: kiro credits
  4. Use available models: kiro ask --model claude-4-5-sonnet "your prompt"

Credit details:

  • Exactly 500 bonus credits for new accounts
  • Credits consumed fractionally (simple prompts use < 1 credit)
  • After 30 days: 50 credits/month on Free tier, or upgrade to paid tiers (1,000-10,000 credits)
  • Credits expire after 30 days if unused
  • Opus 4.5 requires paid tier (no longer on free tier as of Jan 2026)

Sources: kiro.dev/pricing, GitHub issue #5037 Last verified: January 30, 2026

Kilo Code — Opportunistic, not foundational

VOLATILE

What remains useful:

  • Free-model opportunities when the live roster matches what you need
  • BYOK mode if you already have provider keys
  • Tooling flexibility when you want another IDE option

Alternatives for Kimi access:

  1. NVIDIA NIM — Free API trial, no credit card, OpenClaw-compatible ← Best option
  2. OpenCode Zen — Limited-time free models; verify the live roster and data terms
  3. Kimi Code membership — official Kimi coding lane when the live checkout terms fit
  4. Kimi API — pay-as-you-go K3 for newest/1M-context evaluation, K2.7 Code for cheaper coding, and K2.6 only for product or integration-specific compatibility

Tool deep dive: /tools/kilo-code/

Last verified: April 8, 2026

Kimi Code — Membership quota, not a fixed trial promise

What it is: Moonshot AI’s official programming service for Kimi members. Kimi Code now presents itself as a CLI, VS Code extension, and third-party-tool API lane that can expose Kimi K3 for higher-tier users while keeping K2.7 Code as the cheaper coding baseline.

Current official shape:

  • Kimi’s membership page lists Adagio Free plus paid Moderato ($19/mo), Allegretto ($39/mo), Allegro ($99/mo), and Vivace ($199/mo).
  • Kimi Code quota starts on the paid membership rows, from 1x on Moderato up to 30x on Vivace.
  • Current Kimi API docs list Kimi K3 as the newest flagship and Kimi K2.7 Code / Kimi K2.7 Code HighSpeed as the cheaper coding release.
  • The Kimi Code docs use current tool-facing IDs such as k3 and kimi-for-coding; record the displayed backend model because aliases and entitlements can change.
  • Official docs describe five-hour controls, weekly quota refresh, monthly credits, and shared quota across devices and API keys. Treat the signed-in console—not an old fixed request count—as the current limit authority.
  • Membership k3 supports low/high/max reasoning and plan-dependent 256K/1M context; pay-as-you-go kimi-k3 is currently max-only.

API alternative: The Kimi API platform lists Kimi K3 at $0.30 cache-hit input / $3.00 cache-miss input / $15.00 output per 1M tokens. Kimi K2.7 Code remains cheaper at $0.19 / $0.95 / $4.00, with K2.7 Code HighSpeed at $0.38 / $1.90 / $8.00.

Best for: Users who want the official Kimi coding lane in terminal, VS Code, Claude Code-compatible workflows, Roo Code, OpenCode, OpenClaw, or other supported agents.

The catch: Treat trials, haggling, promo banners, and first-month discounts as checkout-controlled. The current official docs support membership tiers and quota mechanics; they do not justify hard-coding an evergreen “$0.99 trial” claim here.

Pricing tiers:

  • Moderato: $19/month (individual, no swarm)
  • Allegretto: $39/month (adds Agent Swarm quota)
  • Allegro: $99/month (larger individual/team quota)
  • Vivace: $199/month (largest listed quota)

Deep dives: /models/kimi-k3/ for the newest flagship, /models/kimi-k2.7-code/ for the cheaper coding lane, and /value/kimi-access/ for Kimi access, membership, and promo-check guidance.

Last verified: July 18, 2026

Optimize Your Spend

Route by Task Complexity

Not everything needs Opus-class pricing:

  • GLM-5.2: 1M context, $1.40/$4.40 API anchor, and a supported-tool Coding Plan path. Worth testing before paying premium API rates for every long coding loop.
  • Kimi K3: latest Kimi flagship with 1M context and Artificial Analysis Index 57. Test it as a Kimi escalation route, not as a cheap default, until CAR is verified.
  • Kimi K2.7 Code: cheaper routine Kimi coding release with cache-hit input, $4.00 output, and a HighSpeed lane when latency matters.
  • Gemini 3 Flash: FREE input tokens, $3/1M output. 78% SWE-bench with 1M context window. Worth testing for high-context workflows, RAG, and document processing before paying Opus or Fable rates.
  • Haiku/Light Models: Route simple requests to cheap models. Save frontier capacity for hard problems.
  • Local Models: Privacy, offline access, zero marginal cost. See /posts/local-llms-guide/

The Compliance Shift

January 2026: Anthropic blocked OpenCode and other third-party clients that spoofed official interfaces.

April 4, 2026: Anthropic formalized the ban — OAuth tokens reserved for Claude Code and Claude.ai only. Third-party harness users must use API keys at full rates.

September 2025: Data retention extended from 30 days to 5 years if you allow training.

What this means: Subscriptions built around third-party tooling are dead. API keys are the only compliant path — and at full rates, the cost gap between Western and Chinese/open-weight providers becomes impossible to ignore.

Verification resources:

Current routing note

Use Sonnet 5 as the first Claude cost/performance test, GLM-5.2 as the cheapest supported-tool lane when quota mechanics fit, and GPT-5.5/ChatGPT Plus for lighter $20 OpenAI use. Keep Kimi K2.7 Code, MiniMax M3, Xiaomi MiMo V2.5, Gemma 4, and Qwen in the comparison table above rather than repeating their canonical pricing pages here. Escalate to Opus 4.8 only when the premium pass changes correctness, architecture, or risk.

Field tests catch failure modes that pricing and benchmark tables flatten. Use How to Read AI Benchmarks to build a shortlist and its mini-eval scorecard to measure cost per accepted result. The LLM app-hacking field test shows why task-level cost and refusal behavior belong in that decision.

Value Section:

Analysis:

Verification:

Comparisons:

July 1 primary sources

Last verified: July 2, 2026. API pricing, subscription multipliers, model-specific limits, regional pricing, and promotional quota rates can change independently.