dot-inference 2.1.0__tar.gz → 2.2.0__tar.gz

{dot_inference-2.1.0 → dot_inference-2.2.0}/CHANGELOG.md

@@ -7,6 +7,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [2.2.0] - 2026-10-08
+
+### Added
+
+- Models: `anthropic/claude-haiku-5.5`, `anthropic/claude-sonnet-5.5`,
+  `anthropic/claude-opus-5.5`, `openai/gpt-6-luna`, `openai/gpt-6.1-sol`,
+  `google/gemini-3.8-flash` and `mistral/mistral-large-4` (public preview;
+  `reasoning_effort` `none` or `high`, the default).
+
 ## [2.1.0] - 2026-10-07
 
 ### Added
@@ -98,7 +107,8 @@ this release.
 - `LLMCallable` and `EmbeddingCallable` return a `Coroutine` rather than the
   looser `Awaitable`, matching what `asyncio.run_coroutine_threadsafe` accepts.
 
-[Unreleased]: https://gitlab.com/deepika6190303/deepika-open-toolbox/dot-inference/-/compare/v2.1.0...main
+[Unreleased]: https://gitlab.com/deepika6190303/deepika-open-toolbox/dot-inference/-/compare/v2.2.0...main
+[2.2.0]: https://gitlab.com/deepika6190303/deepika-open-toolbox/dot-inference/-/tags/v2.2.0
 [2.1.0]: https://gitlab.com/deepika6190303/deepika-open-toolbox/dot-inference/-/tags/v2.1.0
 [2.0.3]: https://gitlab.com/deepika6190303/deepika-open-toolbox/dot-inference/-/tags/v2.0.3
 [2.0.2]: https://gitlab.com/deepika6190303/deepika-open-toolbox/dot-inference/-/tags/v2.0.2

{dot_inference-2.1.0 → dot_inference-2.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: dot-inference
-Version: 2.1.0
+Version: 2.2.0
 Summary: Unified async Python library for LLM inference and embeddings across multiple providers
 Project-URL: Homepage, https://gitlab.com/deepika6190303/deepika-open-toolbox/dot-inference
 Project-URL: Repository, https://gitlab.com/deepika6190303/deepika-open-toolbox/dot-inference
@@ -332,6 +332,7 @@ See [docs/VERSIONING.md](docs/VERSIONING.md) for the full policy.
 | [docs/DEVELOPMENT.md](docs/DEVELOPMENT.md) | Environment setup, tests, code style |
 | [docs/VERSIONING.md](docs/VERSIONING.md) | Versioning, deprecation policy, how to depend on this package |
 | [docs/PUBLISHING.md](docs/PUBLISHING.md) | Cutting a release |
+| [docs/provider-quirks/](docs/provider-quirks/) | Provider behaviours that differ from the OpenAI format: `tool_choice` support, OpenRouter error bodies |
 | [CHANGELOG.md](CHANGELOG.md) | Release history |
 
 ## Contributing

{dot_inference-2.1.0 → dot_inference-2.2.0}/README.md

@@ -305,6 +305,7 @@ See [docs/VERSIONING.md](docs/VERSIONING.md) for the full policy.
 | [docs/DEVELOPMENT.md](docs/DEVELOPMENT.md) | Environment setup, tests, code style |
 | [docs/VERSIONING.md](docs/VERSIONING.md) | Versioning, deprecation policy, how to depend on this package |
 | [docs/PUBLISHING.md](docs/PUBLISHING.md) | Cutting a release |
+| [docs/provider-quirks/](docs/provider-quirks/) | Provider behaviours that differ from the OpenAI format: `tool_choice` support, OpenRouter error bodies |
 | [CHANGELOG.md](CHANGELOG.md) | Release history |
 
 ## Contributing

dot_inference-2.2.0/docs/provider-quirks/openrouter-errors.md

@@ -0,0 +1,19 @@
+# OpenRouter error bodies
+
+OpenRouter translates requests into each provider's format, but passes the
+provider's error back untranslated, wrapped in a generic envelope
+([Errors and debugging](https://openrouter.ai/docs/api/reference/errors-and-debugging#provider-errors)):
+
+```json
+{"error": {"code": 400, "message": "Provider returned error",
+           "metadata": {"raw": "<the provider's error body, verbatim>",
+                        "provider_name": "Anthropic"}}}
+```
+
+- `message` is generic; the actual cause is in `metadata.raw`.
+- `metadata.raw` uses the provider's vocabulary, not the OpenAI one used in
+  the request. Example: a `tool_choice: "required"` refused by Anthropic is
+  reported as `type "tool" and "any" are not supported` (see
+  [tool-choice.md](tool-choice.md)).
+- `metadata.provider_name` names the provider that failed — useful when
+  routing is not pinned with `provider_routing`.

dot_inference-2.2.0/docs/provider-quirks/tool-choice.md

@@ -0,0 +1,31 @@
+# `tool_choice` across providers
+
+dot-inference sends `tool_choice` in OpenAI format. Its meaning
+([OpenAI](https://platform.openai.com/docs/guides/function-calling),
+[OpenRouter](https://openrouter.ai/docs/guides/features/tool-calling#tool-choice-configuration)):
+
+| Value | Meaning |
+|---|---|
+| `"auto"` | The model calls tools or answers in text, as it decides. Default when `tools` is given. |
+| `"none"` | Tools stay in the request but the model must answer in text. Keeps the prompt (and its cache) identical to a tool-calling turn — e.g. a summarisation step. |
+| `"required"` | The model must call at least one tool. |
+| `{"type": "function", "function": {"name": "X"}}` | The model must call tool `X`. |
+
+## Anthropic models via OpenRouter
+
+OpenRouter translates the request into Anthropic's format
+([Anthropic `tool_choice`](https://platform.claude.com/docs/en/agents-and-tools/tool-use/define-tools#forcing-tool-use)):
+`"required"` becomes `{"type": "any"}` and a named function becomes
+`{"type": "tool", "name": "X"}`.
+
+Anthropic does not support forced tool use (`any`, `tool`) on **Claude Opus
+5.5, Sonnet 5.5, Fable 5.1**, nor with manual extended thinking. These calls
+fail with a 400 — `auto` and `none` still work. Haiku 5.5 and the 4.x models
+accept it. Observed 2026-10-08 through OpenRouter:
+
+```
+tool_choice: type "tool" and "any" are not supported for this model.
+```
+
+The message names Anthropic's types, not the `"required"` that was sent: see
+[openrouter-errors.md](openrouter-errors.md).

{dot_inference-2.1.0 → dot_inference-2.2.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "dot-inference"
-version = "2.1.0"
+version = "2.2.0"
 description = "Unified async Python library for LLM inference and embeddings across multiple providers"
 readme = "README.md"
 authors = [{ name = "Kannon For Deep Tech", email = "louis.letarnec@deepika.ai" }]

{dot_inference-2.1.0 → dot_inference-2.2.0}/src/dot_inference/models.yaml

@@ -221,6 +221,39 @@ openai/gpt-5.4:
         provider_routing:
           only: ["openai"]
 
+# GPT-6 Luna (September 2026): the fast, low-cost tier of the GPT-6 series.
+openai/gpt-6-luna:
+  supports_strict_json: true
+  supports_reasoning: true
+  input_token_cost: 1e-7      # $0.10/M
+  output_token_cost: 5e-7     # $0.50/M
+  context_window: 1050000
+  providers:
+    openai:
+      api_model_name: gpt-6-luna
+    openrouter:
+      api_model_name: openai/gpt-6-luna
+      features:
+        provider_routing:
+          only: ["openai"]
+
+# GPT-6.1 Sol (September 2026): the mid tier of the GPT-6 series, below
+# GPT-6 Astra; an upgrade of GPT-6 Sol at the same price.
+openai/gpt-6.1-sol:
+  supports_strict_json: true
+  supports_reasoning: true
+  input_token_cost: 2e-6      # $2/M
+  output_token_cost: 1e-5     # $10/M
+  context_window: 1050000
+  providers:
+    openai:
+      api_model_name: gpt-6.1-sol
+    openrouter:
+      api_model_name: openai/gpt-6.1-sol
+      features:
+        provider_routing:
+          only: ["openai"]
+
 openai/gpt-oss-120b:
   supports_strict_json: true
   supports_reasoning: true
@@ -273,6 +306,17 @@ google/gemini-3-pro:
     openrouter:
       api_model_name: google/gemini-3.1-pro-preview
 
+# Gemini 3.8 Flash (September 2026).
+google/gemini-3.8-flash:
+  supports_strict_json: true
+  supports_reasoning: true
+  input_token_cost: 7.5e-7    # $0.75/M
+  output_token_cost: 3.75e-6  # $3.75/M
+  context_window: 1048576
+  providers:
+    openrouter:
+      api_model_name: google/gemini-3.8-flash
+
 # Gemma 4 31B (v26.04, April 2026): open-weight (Apache 2.0) dense multimodal
 # model with native function calling, structured outputs, and a configurable
 # reasoning mode. Served across providers on OpenRouter.
@@ -305,6 +349,29 @@ mistral/mistral-large-latest:
         provider_routing:
           only: ["mistral"]
 
+# Mistral Large 4 (v26.10, public preview since October 2026): open-weight
+# multimodal MoE (1.05T parameters). Pinned ID: on La Plateforme
+# mistral-large-latest still resolves to Large 3 (mistral-large-2512).
+mistral/mistral-large-4:
+  default_provider: mistral
+  supports_strict_json: true
+  # Adjustable reasoning, same contract as Medium 3.5: La Plateforme accepts
+  # only "none" and "high", and thinks ("high") when none is given.
+  supports_reasoning: true
+  reasoning_effort_options: [none, high]
+  default_reasoning_effort: high
+  input_token_cost: 6.8e-7    # $0.68/M (preview price; list $1.36/M)
+  output_token_cost: 2.09e-6  # $2.09/M (preview price; list $4.18/M)
+  context_window: 524288
+  providers:
+    mistral:
+      api_model_name: mistral-large-4
+    openrouter:
+      api_model_name: mistralai/mistral-large-4-0
+      features:
+        provider_routing:
+          only: ["mistral"]
+
 
 # Alias — always resolves to the latest Mistral Small release.
 # Use this when you want automatic upgrades as Mistral ships new versions.
@@ -483,3 +550,49 @@ anthropic/claude-opus-4.6:
       features:
         provider_routing:
           only: ["anthropic"]
+
+# Claude Haiku 5.5 (October 2026). Priced by prompt length: the costs below
+# apply up to 100K prompt tokens; above, Anthropic bills $0.50/M input and
+# $2.50/M output. OpenRouter reports the billed cost, which takes precedence
+# over this estimate.
+anthropic/claude-haiku-5.5:
+  supports_strict_json: true
+  supports_reasoning: true
+  requires_cache_control: true
+  input_token_cost: 1e-7      # $0.10/M
+  output_token_cost: 5e-7     # $0.50/M
+  context_window: 1000000
+  providers:
+    openrouter:
+      api_model_name: anthropic/claude-haiku-5.5
+      features:
+        provider_routing:
+          only: ["anthropic"]
+
+anthropic/claude-sonnet-5.5:
+  supports_strict_json: true
+  supports_reasoning: true
+  requires_cache_control: true
+  input_token_cost: 2e-6      # $2/M
+  output_token_cost: 1e-5     # $10/M
+  context_window: 1000000
+  providers:
+    openrouter:
+      api_model_name: anthropic/claude-sonnet-5.5
+      features:
+        provider_routing:
+          only: ["anthropic"]
+
+anthropic/claude-opus-5.5:
+  supports_strict_json: true
+  supports_reasoning: true
+  requires_cache_control: true
+  input_token_cost: 4e-6      # $4/M
+  output_token_cost: 2e-5     # $20/M
+  context_window: 1000000
+  providers:
+    openrouter:
+      api_model_name: anthropic/claude-opus-5.5
+      features:
+        provider_routing:
+          only: ["anthropic"]

{dot_inference-2.1.0 → dot_inference-2.2.0}/uv.lock

@@ -119,7 +119,7 @@ wheels = [
 
 [[package]]
 name = "dot-inference"
-version = "2.1.0"
+version = "2.2.0"
 source = { editable = "." }
 dependencies = [
     { name = "httpx" },