Changelog

Notable changes to the Nexotao API and model catalog. Breaking changes and billing-affecting changes are called out explicitly.

Scheduled 2026-08-05 00:00 WIB — model rate adjustment

Rates for all text models are scheduled to change on Wednesday, 5 August 2026 at 00:00 WIB (UTC+7). See the new rate table and live countdown.

The new rates are set 25% below official provider prices. Current Rupiah figures remain available on the pricing page and through GET /models.

• Existing rates remain in effect through 4 August at 23:59 WIB.
• New rates apply only to requests after the effective time. Completed usage is not rebilled.
• The nominal amount of your Rupiah balance does not change. After cutover, each request is deducted at the rate published by GET /models when it runs.
• Claude and GPT cache tokens keep the catalog policy: cache reads at 10% and cache writes at 125% of the input rate then in effect.

During cutover, some models may temporarily fail or respond more slowly than usual. Watch the dashboard banner and retry with backoff; other services remain operational.

2026-07-26 — Claude Opus 5 is live, with a 1M-token context

claude-opus-5 joins the catalog through the same Anthropic endpoint. Change only the model value; your API key, base URL, and Rupiah balance stay the same.

• Introductory pricing below the standing Opus tier, running until Tuesday, 28 July 2026 at 23:59 WIB (UTC+7).

  Update (2026-07-30): the introductory period has ended. The price increase announced in this entry was not applied — Opus 5 still bills at its introductory rate. Current rates are always in GET /models and on Models & Pricing.

• 1,000,000-token context window — far above the 350k of the Opus 4.x line. Read context_window from /v1/models; don’t carry one model’s window over to another. See Context Window & Compaction.

• Vision is supported (image blocks on /v1/messages).

• Sampling parameters are rejected. Like Opus 4.8 and 4.7, this model returns 400 for temperature, top_p, and top_k. The gateway strips all three before forwarding, so existing clients keep working — but the values you send have no effect.

• Capacity. Opus 5 is served by seven upstream deployments, so there is failover when one is saturated. Concurrent requests per model are still capped so no single deployment saturates, so a sharp traffic spike can return 429 server at capacity, retry shortly — handle it the usual way, with retry + backoff. The cap will be raised as real traffic patterns emerge.

2026-07-15 — GPT-5.6 Sol, Terra, and Luna are live

Three new GPT-5.6 deployments are available through the same OpenAI-compatible endpoint. Change only the model value; your API key, base URL, and Rupiah balance stay the same.

• gpt-5.6-luna — Rp 2,349 input / Rp 4,698 output per 1M tokens.
• gpt-5.6-terra — Rp 4,500 input / Rp 9,000 output per 1M tokens.
• gpt-5.6-sol — Rp 7,000 input / Rp 7,000 output per 1M tokens.

GPT cache tokens use the verified cache rates: reads at 10% of input and writes at 125% of input. The authoritative model list and rates are always available from GET /models and Models & Pricing.

2026-07-13 — Clearer model positioning: DeepSeek for volume/cheap, Claude for agentic

The docs now spell out model selection. DeepSeek-V4-Flash is for high-volume, cheap work (classification, extraction, short chat); Claude (Opus/Sonnet) is for coding/agentic & large context. DeepSeek has a small capacity (~20k tokens/min), so a single agentic or large-context request can return 429 server at capacity, retry shortly. See the FAQ. No API or pricing change — guidance alignment only.

2026-06-26 — Opus 4.8 rate update (+30%)

Claude Opus 4.8 (claude-opus-4-8) received a new published price on this date, a +30% change from the previous rate.

Update (2026-07-30): the figures in this entry are a historical record as of 2026-06-26 and are not the current rate — Opus 4.8 has been repriced several times since. Current rates are always in GET /models and on Models & Pricing.

What changed at the time

• The Opus 4.8 rate rose by +30% from the previous rate.
• DeepSeek-V4-Pro is unchanged.
• Your balance keeps full value — no credits are lost or revalued, and past usage is not re-billed; the new rate applies to usage from here forward.

Why it matters

You bill in Rupiah at the rate published when the request runs. A rate change applies to usage going forward — usage already settled at the old rate is not re-billed. If your cheaper-first workloads run on DeepSeek-V4-Pro, nothing changes for them.

2026-06-25 — Published context window: read the real number, no silent truncation

Every live model now publishes a verified context_window, so your client calibrates to the true window instead of hard-guessing 200k — and we never silently drop tokens behind your back.

What’s new

• context_window is now published per model on both GET /models (no key) and GET /v1/models. It’s a token count, and it’s the source of truth — always in sync with what the gateway actually serves.
• Current published windows:
  • claude-opus-4-8 → 350,000 tokens — larger than the legacy 200k many clients still assume by default.
  • DeepSeek-V4-Pro → 131,072 tokens.

Note (2026-07-30): context_window is only published for models whose hard limit has been verified; for the others the field is null. The current list is on Context Window & Compaction.

Why it matters

Some endpoints silently truncate your context when a conversation outgrows the window: the agent keeps running but quietly loses the earliest messages, and you only find out when the answers get worse. We don’t do that. When a prompt genuinely exceeds the window you get an explicit error — not a silent drop. And because we publish the real number, a calibrated client (for example Claude Code) compacts at the right time instead of tripping early against a guessed 200k.

# Read the real window — no key needed
curl https://api.nexotao.com/models   # each model carries context_window

What you should do

1. Read context_window from the API instead of hard-coding 200k (or assuming 1M — the published Opus window is 350k, not more).
2. Let your client compact. See the Context Window & Compaction guide to calibrate Claude Code and keep long sessions going.

2026-06-24 — Catalog trim: two models, deprecated ids on a 30-day grace window

We’ve trimmed the live model catalog to two models. If your code references any other model id, read the billing note below — it can affect what you’re charged during the grace window.

What changed

Update (2026-07-18): this catalog trim was later reversed. The models claude-sonnet-4-6, gpt-5-mini, and DeepSeek-V4-Flash are active again as live, independently-priced models and were not removed. The entry below is kept as a historical record. See Models & Pricing for the current live catalog.

• The live catalog is now two chat models only:

  • claude-opus-4-8 (Claude Opus 4.8)
  • DeepSeek-V4-Pro (DeepSeek V4 Pro)

• Image generation and audio transcription are no longer available. POST /v1/images/generations and POST /v1/audio/transcriptions are no longer served — there are no active image or transcription models in the catalog.

  Correction (2026-07-30): this entry originally said both endpoints return 400. In fact the routes are not registered at all, so both return 404 with the code route.not_found.

• Deprecated chat ids still work during a grace window, then stop. A request for a deprecated id is transparently routed to a live model during the grace window. Deprecated ids are planned for removal on or around 2026-07-24, after which they return 400 unknown model. We’ll announce the exact cutover.

Billing note — please read

During the grace window, a deprecated id is billed at the live model’s rate, not the old id’s rate. Several deprecated ids were cheaper than the model they now route to, so a request that used to bill at a lower rate will bill at the canonical (higher) rate.

This is balance-bounded — you are never billed for more than your balance covers, and a request that can’t be covered is rejected rather than overdrawing — but it is a real price change for those ids. To avoid surprises, update your code to call a live id directly now rather than waiting for the cutoff.

Deprecated id → live model

Retired media models (no replacement): FLUX.2-pro, gpt-4o-transcribe, gpt-4o-mini-transcribe, whisper.

What you should do

1. Update your model id to claude-opus-4-8 or DeepSeek-V4-Pro.
2. Check your billing assumptions if you were using a deprecated cheaper id — you’re now billed at the live model’s rate.
3. Migrate soon. Deprecated ids are planned for removal on or around 2026-07-24, after which they return 400.

Live rates are always available from GET /models and on the Models & Pricing page. Questions? See the FAQ.