v0.9.1

Refresh pricing data to 318 models — adds grok-4.3, grok-build-0.1, claude-opus-4.8, o3-pro, gpt-4.1, o4-mini, and more.

v0.9.0 — Claude 4.7, GPT-5.5, pricing update May 2026

New models

Anthropic

• claude-opus-4-7 — Most capable model, step-change in agentic coding, 1M context, 128K max output ($5/$25 per 1M tokens)
• claude-opus-4-5 — Improved coding + workplace tasks ($5/$25)
• claude-opus-4-1 — Legacy frontier ($15/$75)
• Updated context lengths for Claude 4.6+ to 1M tokens

OpenAI

• gpt-5.5 — Smart/coding flagship, 1M context ($5/$30 per 1M tokens)
• gpt-5.5-pro — Top flagship, 1M context ($30/$180 per 1M tokens)

Pricing

Live refresh via OpenRouter API — 338 models total (was 321). Includes deepseek-v4, claude-opus-4.6-fast, and more.

Fixes

• Pricing lookup now correctly matches claude-opus-4-7 (Anthropic API dash format) to claude-opus-4.7 (OpenRouter dot format)
• fetch-pricing.ts now preserves the dot/dash normalization fix on future runs
• Added o5 prefix to OpenAI model listing filter

v0.8.1

README improvements: added badges, star ask, and documentation updates.

v0.8.0 — Adaptive Concurrency & max_tokens Translation

What's New

Adaptive Concurrency for Concurrent Batches

Set concurrencyFallback: 'auto' to dynamically discover your provider's rate limit ceiling. Uses TCP-style slow-start (5 → 10 → 20 → 40 → ...) then AIMD for fine-tuning. Reads x-ratelimit-remaining-requests headers proactively — an OpenAI Tier 4 account ramps to ~160 concurrent in ~155 requests instead of being stuck at 5.

Add concurrencyMax for a hard ceiling when multiple jobs share an API key.

max_tokens → max_completion_tokens Translation

Newer OpenAI models (gpt-4o, o1, o3, o4-mini, gpt-5-mini) require max_completion_tokens. anymodel now detects the model and sends the correct parameter automatically — callers always just pass max_tokens.

Fixes

• Flex discount now applied on concurrent batch cost calculations (was only applied for native batches)
• service_tier persisted on BatchObject for accurate cost retrieval
• gpt-5-mini added to token estimation limits
• RateLimitTracker now wired up — provider response headers feed into fallback routing decisions

Full changelog: https://github.com/probeo-io/anymodel/blob/main/CHANGELOG.md

v0.7.1

Fixed

• Batch cost calculations now apply 50% discount for native batch APIs (OpenAI, Anthropic, Google)
• BatchBuilder accounts for flex pricing discount on concurrent batches

v0.7.0

Added

• BatchBuilder API — ergonomic batch construction with client.batches.open(config)
  • add(prompt) — just pass strings, persisted to disk immediately
  • submit() — formats for provider (Anthropic/OpenAI/Google) and dispatches
  • poll() — returns clean succeeded/failed results with per-item costs
  • retry(failed) — creates a new builder pre-loaded with failed items
  • No custom_ids, no message formatting, no provider-specific formats — all hidden
• Poll logging: logToConsole option and ANYMODEL_BATCH_POLL_LOG env var

v0.6.0

Added

• Automatic per-request cost calculation from bundled pricing data (323 models)
• Pricing fetched from OpenRouter at build time — always current as of last publish
• GenerationStats.total_cost now calculated automatically from token usage
• BatchUsageSummary.estimated_cost now calculated automatically from token usage
• Exported getModelPricing(), calculateCost(), PRICING_AS_OF, PRICING_MODEL_COUNT
• batch_mode option on BatchCreateRequest — set to "concurrent" to force individual requests (e.g. for flex pricing)
• service_tier support on batch requests (concurrent path only — native batch already discounted)

v0.5.3

Added

• batch_mode option on BatchCreateRequest — set to "concurrent" to force individual requests instead of native batch API (e.g. to use flex pricing on OpenAI). Auto-detects when omitted (no breaking change).

v0.5.2

Added

• service_tier support on batch requests — set per-request or as a batch-level default
• Concurrent batch passes service_tier through to chat completions for flex pricing
• Native OpenAI batch intentionally omits service_tier (already gets 50% off)

v0.5.1

Changed

• Concurrent batch processing now streams requests from disk instead of holding all in memory — safe for 10K+ request batches
• Only N requests (default 5) are in-flight at a time, the rest stay on disk until needed