python-fastllm 0.0.33__tar.gz → 0.0.35__tar.gz

{python_fastllm-0.0.33 → python_fastllm-0.0.35}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-fastllm
-Version: 0.0.33
+Version: 0.0.35
 Author-email: Kerem Turgutlu <keremturgutlu@gmail.com>
 License: Apache-2.0
 Project-URL: Repository, https://github.com/AnswerDotAI/fastllm
@@ -10,7 +10,7 @@ Classifier: Programming Language :: Python :: 3 :: Only
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 Requires-Dist: fastcore>=2.1.11
-Requires-Dist: aidialog>=0.0.3
+Requires-Dist: aidialog>=0.0.5
 Requires-Dist: fastspec>=0.0.11
 Requires-Dist: toolslm
 Requires-Dist: pillow

python_fastllm-0.0.35/fastllm/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.0.35"

{python_fastllm-0.0.33 → python_fastllm-0.0.35}/fastllm/_modidx.py

@@ -10,6 +10,7 @@ d = { 'settings': { 'branch': 'main',
                                    'fastllm.acomplete._classify_error': ('acomplete.html#_classify_error', 'fastllm/acomplete.py'),
                                    'fastllm.acomplete._classify_error_stream': ( 'acomplete.html#_classify_error_stream',
                                                                                  'fastllm/acomplete.py'),
+                                   'fastllm.acomplete._codex_mtime_policy': ('acomplete.html#_codex_mtime_policy', 'fastllm/acomplete.py'),
                                    'fastllm.acomplete._debug_print': ('acomplete.html#_debug_print', 'fastllm/acomplete.py'),
                                    'fastllm.acomplete._is_ctx_exceeded': ('acomplete.html#_is_ctx_exceeded', 'fastllm/acomplete.py'),
                                    'fastllm.acomplete._raise_if_done': ('acomplete.html#_raise_if_done', 'fastllm/acomplete.py'),
@@ -94,7 +95,6 @@ d = { 'settings': { 'branch': 'main',
                               'fastllm.chat.ToolReminderCallback.__init__': ('chat.html#toolremindercallback.__init__', 'fastllm/chat.py'),
                               'fastllm.chat.ToolReminderCallback.after_msgs': ( 'chat.html#toolremindercallback.after_msgs',
                                                                                 'fastllm/chat.py'),
-                              'fastllm.chat.ToolResponse': ('chat.html#toolresponse', 'fastllm/chat.py'),
                               'fastllm.chat.UsageStats': ('chat.html#usagestats', 'fastllm/chat.py'),
                               'fastllm.chat.UsageStats.__add__': ('chat.html#usagestats.__add__', 'fastllm/chat.py'),
                               'fastllm.chat.UsageStats.__init__': ('chat.html#usagestats.__init__', 'fastllm/chat.py'),
@@ -106,7 +106,6 @@ d = { 'settings': { 'branch': 'main',
                               'fastllm.chat._add_cache_control': ('chat.html#_add_cache_control', 'fastllm/chat.py'),
                               'fastllm.chat._alite_call_func': ('chat.html#_alite_call_func', 'fastllm/chat.py'),
                               'fastllm.chat._apply_cache_idxs': ('chat.html#_apply_cache_idxs', 'fastllm/chat.py'),
-                              'fastllm.chat._bytes2content': ('chat.html#_bytes2content', 'fastllm/chat.py'),
                               'fastllm.chat._call_func': ('chat.html#_call_func', 'fastllm/chat.py'),
                               'fastllm.chat._handle_stop_reason': ('chat.html#_handle_stop_reason', 'fastllm/chat.py'),
                               'fastllm.chat._has_cache': ('chat.html#_has_cache', 'fastllm/chat.py'),
@@ -114,14 +113,11 @@ d = { 'settings': { 'branch': 'main',
                               'fastllm.chat._has_stop': ('chat.html#_has_stop', 'fastllm/chat.py'),
                               'fastllm.chat._inject_tool_reminder': ('chat.html#_inject_tool_reminder', 'fastllm/chat.py'),
                               'fastllm.chat._lite_call_func': ('chat.html#_lite_call_func', 'fastllm/chat.py'),
-                              'fastllm.chat._mime2part_type': ('chat.html#_mime2part_type', 'fastllm/chat.py'),
-                              'fastllm.chat._mk_content': ('chat.html#_mk_content', 'fastllm/chat.py'),
                               'fastllm.chat._mk_prefill': ('chat.html#_mk_prefill', 'fastllm/chat.py'),
                               'fastllm.chat._mk_tool_result': ('chat.html#_mk_tool_result', 'fastllm/chat.py'),
                               'fastllm.chat._tc_summary': ('chat.html#_tc_summary', 'fastllm/chat.py'),
                               'fastllm.chat._think_kw': ('chat.html#_think_kw', 'fastllm/chat.py'),
                               'fastllm.chat._trunc_param': ('chat.html#_trunc_param', 'fastllm/chat.py'),
-                              'fastllm.chat._url2content': ('chat.html#_url2content', 'fastllm/chat.py'),
                               'fastllm.chat._usrtools': ('chat.html#_usrtools', 'fastllm/chat.py'),
                               'fastllm.chat.add_warning': ('chat.html#add_warning', 'fastllm/chat.py'),
                               'fastllm.chat.adisplay_stream': ('chat.html#adisplay_stream', 'fastllm/chat.py'),
@@ -253,12 +249,15 @@ d = { 'settings': { 'branch': 'main',
                                    'fastllm.streaming.norm_and_yield': ('streaming.html#norm_and_yield', 'fastllm/streaming.py'),
                                    'fastllm.streaming.stop_and_trim': ('streaming.html#stop_and_trim', 'fastllm/streaming.py')},
             'fastllm.types': { 'fastllm.types.APIRegistry': ('types.html#apiregistry', 'fastllm/types.py'),
-                               'fastllm.types.APIRegistry.__init__': ('types.html#apiregistry.__init__', 'fastllm/types.py'),
+                               'fastllm.types.APIRegistry.__contains__': ('types.html#apiregistry.__contains__', 'fastllm/types.py'),
+                               'fastllm.types.APIRegistry.__missing__': ('types.html#apiregistry.__missing__', 'fastllm/types.py'),
+                               'fastllm.types.APIRegistry._load': ('types.html#apiregistry._load', 'fastllm/types.py'),
                                'fastllm.types.APIRegistry.register': ('types.html#apiregistry.register', 'fastllm/types.py'),
                                'fastllm.types.Completion': ('types.html#completion', 'fastllm/types.py'),
                                'fastllm.types.Completion._repr_markdown_': ('types.html#completion._repr_markdown_', 'fastllm/types.py'),
                                'fastllm.types.Completion.cost': ('types.html#completion.cost', 'fastllm/types.py'),
                                'fastllm.types.Usage': ('types.html#usage', 'fastllm/types.py'),
+                               'fastllm.types._rm_ctx_tiers': ('types.html#_rm_ctx_tiers', 'fastllm/types.py'),
                                'fastllm.types.approx_pricing': ('types.html#approx_pricing', 'fastllm/types.py'),
                                'fastllm.types.fn_schema': ('types.html#fn_schema', 'fastllm/types.py'),
                                'fastllm.types.get_api_key': ('types.html#get_api_key', 'fastllm/types.py'),
@@ -270,5 +269,7 @@ d = { 'settings': { 'branch': 'main',
                                'fastllm.types.mk_completion': ('types.html#mk_completion', 'fastllm/types.py'),
                                'fastllm.types.model_prices_meta': ('types.html#model_prices_meta', 'fastllm/types.py'),
                                'fastllm.types.payload_kwargs': ('types.html#payload_kwargs', 'fastllm/types.py'),
+                               'fastllm.types.price_tier': ('types.html#price_tier', 'fastllm/types.py'),
                                'fastllm.types.register_model_info': ('types.html#register_model_info', 'fastllm/types.py'),
-                               'fastllm.types.resize_b64': ('types.html#resize_b64', 'fastllm/types.py')}}}
+                               'fastllm.types.resize_b64': ('types.html#resize_b64', 'fastllm/types.py'),
+                               'fastllm.types.tier_rate': ('types.html#tier_rate', 'fastllm/types.py')}}}

{python_fastllm-0.0.33 → python_fastllm-0.0.35}/fastllm/acomplete.py

@@ -54,14 +54,29 @@ api2spec = {'openai':oai_spec, 'openai_chat':oai_spec, 'anthropic':ant_spec, 'ge
 
 # %% ../nbs/06_acomplete.ipynb #e3ed40fb
 def split_vendor(model):
-    "Split a `'vendor/model'` path into `(vendor_name, model)`: the prefix must name a known vendor or a registered api, else `(None, model)`."
+    "Split a `'vendor/model'` path into `(vendor_name, model)`: the prefix must name a known vendor or a registered api (installed `fastllm.apis` plugins load on demand), else `(None, model)`."
     if model and '/' in model:
         v, m = model.split('/', 1)
-        if v in vendor_mapping or v in api_registry.apis: return v, m
+        if v in vendor_mapping or v in api_registry: return v, m
     return None, model
 
+# %% ../nbs/06_acomplete.ipynb #977a115b
+def _codex_mtime_policy():
+    "`flexicache` policy: expire codex clients when the codex auth file changes"
+    def policy(state, args, kwargs):
+        v = kwargs.get('vendor_name') or (args[1] if len(args)>1 else None)
+        if not v:
+            m = kwargs.get('model') or (args[0] if args else None)
+            v = split_vendor(m)[0] if m else None
+        if v != 'codex': return None
+        try: mt = Path(_codex_path).expanduser().stat().st_mtime
+        except OSError: mt = None
+        return mt if state != mt else None
+    policy.needs_args = True
+    return policy
+
 # %% ../nbs/06_acomplete.ipynb #79075d95
-@flexicache()
+@flexicache(_codex_mtime_policy())
 def mk_client(model=None, vendor_name=None, api_name=None, api_key=None, base_url=None, xtra_hdrs=None,
     timeout=httpx.Timeout(connect=30, read=300, write=30, pool=10)):
     err_msg = f"please pass a valid one vendor: {', '.join(list(vendor_mapping))} or pass `api_name`,`base_url` and `api_key`"
@@ -82,7 +97,7 @@ def mk_client(model=None, vendor_name=None, api_name=None, api_key=None, base_ur
     elif base_url and api_key: vendor_name, api_name = ifnone(vendor_name, 'custom'), ifnone(api_name, 'openai_chat')
     elif (api_name:=infer_api_name(model)):  base_url, vendor_name = None, api_name
     else: raise ValueError(f"Model {model} can't be auto resolved, {err_msg}")
-    api = api_registry.apis[api_name]
+    api = api_registry[api_name]
     spec, hdrs = api2spec[api_name], api.get_hdrs(api_key)
     cli = OpenAPIClient(spec, headers=merge(hdrs, ifnone(xtra_hdrs, {})), timeout=timeout)
     if base_url is not None:
@@ -163,13 +178,13 @@ async def acomplete(msgs, model, api_name=None, vendor_name=None, api_key=None,
         elif v: api_name, model = v, m  # a registered transport api (e.g. claude_code): not an HTTP vendor
     if api_name == 'claude_code':
         if not stream: raise NotImplementedError("claude_code backend supports stream=True only for now")
-        api = api_registry.apis[api_name]
+        api = api_registry[api_name]
         payload = api.mk_payload(msgs, model, stream=stream, **kwargs)
         async def _mk_gen():
             async for o in api.acollect_stream(payload, model=model, vendor_name='claude_code', stop_callables=stop_callables): yield o
         return _retry_stream(_mk_gen, retries, retry_delay)
-    cli, api_name, vendor_name = mk_client(model, vendor_name, api_name, api_key, base_url, xtra_hdrs)
-    api = api_registry.apis[api_name]
+    cli, api_name, vendor_name = mk_client(model=model, vendor_name=vendor_name, api_name=api_name, api_key=api_key, base_url=base_url, xtra_hdrs=xtra_hdrs)
+    api = api_registry[api_name]
     payload = api.mk_payload(msgs, model, stream=stream, **kwargs)
     payload = merge(payload, ifnone(xtra_body, {}))
     if vendor_name == 'codex':

{python_fastllm-0.0.33 → python_fastllm-0.0.35}/fastllm/chat.py

@@ -4,10 +4,10 @@
 
 # %% auto #0
 __all__ = ['effort', 'remove_cache_ckpts', 'contents', 'stop_reason', 'mk_msg', 'FenceToolStop', 'extract_fence_call', 'mk_msgs',
-           'cite_footnote', 'postproc', 'lite_mk_func', 'ToolResponse', 'structured', 'search_count', 'UsageStats',
-           'AsyncChat', 'astream_with_complete', 'ChatCallback', 'DeepseekMsgsCallback', 'DeepseekPrefillCallback',
-           'add_warning', 'StopReasonCallback', 'run_fence_tool', 'FenceToolCallback', 'ToolReminderCallback',
-           'stop_sequences', 'StopSequencesCallback', 'StreamFormatter', 'AsyncStreamFormatter', 'adisplay_stream']
+           'cite_footnote', 'postproc', 'lite_mk_func', 'structured', 'search_count', 'UsageStats', 'AsyncChat',
+           'astream_with_complete', 'ChatCallback', 'DeepseekMsgsCallback', 'DeepseekPrefillCallback', 'add_warning',
+           'StopReasonCallback', 'run_fence_tool', 'FenceToolCallback', 'ToolReminderCallback', 'stop_sequences',
+           'StopSequencesCallback', 'StreamFormatter', 'AsyncStreamFormatter', 'adisplay_stream']
 
 # %% ../nbs/07_chat.ipynb #d5a3bc1f
 from typing import Optional,Callable
@@ -15,35 +15,13 @@ from toolslm.funccall import mk_ns, call_func, call_func_async, get_schema
 from fastcore.utils import *
 from fastcore.meta import delegates
 from dataclasses import dataclass
-import base64
 
 from .types import *
-from fastcore.xtras import detect_mime
 from aidialog.msg_parts import (Msg, Part, PartType, ToolCall, mk_tool_res_msg, mk_tr_details,
-    tool_info, usage_info, think_start, think_end, StopResponse, display_list, MediaUrl, url_mime, fmt2hist,
-    _mk_result_fence, _split_fence_msgs, _fence_back, _trunc_str)
+    tool_info, usage_info, think_start, think_end, StopResponse, ToolResponse, display_list, fmt2hist,
+    _mk_content, _mk_result_fence, _split_fence_msgs, _fence_back, _trunc_str)
 from .acomplete import *
 
-# %% ../nbs/07_chat.ipynb #4a37bae5
-def _mime2part_type(mime):
-    "Map MIME string to canonical PartType"
-    if mime.startswith('image/'): return PartType.input_image
-    if mime.startswith('audio/'): return PartType.input_audio
-    if mime.startswith('video/'): return PartType.input_video
-    return PartType.input_file
-
-def _bytes2content(data):
-    "Convert bytes to fastllm canonical content"
-    mtype = detect_mime(data)
-    if not mtype: raise ValueError(f'Data must be a supported file type, got {data[:10]}')
-    encoded = base64.b64encode(data).decode("utf-8")
-    return Part(type=_mime2part_type(mtype), text=f'data:{mtype};base64,{encoded}')
-
-def _url2content(o):
-    "Convert MediaUrl to fastllm canonical content"
-    mime = o.mime or url_mime(o.url)
-    return Part(type=_mime2part_type(mime), text=o.url, data=dict(mime=mime))
-
 # %% ../nbs/07_chat.ipynb #43a55cda
 def _add_cache_control(
     msg,          # LiteLLM formatted msg
@@ -67,12 +45,6 @@ def remove_cache_ckpts(msg):
         if part.data: part.data.pop('cache_control', None)
     return msg
 
-def _mk_content(o):
-    if isinstance(o, str):        return Part(type=PartType.text, text=o)
-    elif isinstance(o, bytes):    return _bytes2content(o)
-    elif isinstance(o, MediaUrl): return _url2content(o)
-    return o
-
 
 # %% ../nbs/07_chat.ipynb #48c78e48
 def contents(c):
@@ -169,11 +141,6 @@ def lite_mk_func(f):
     if isinstance(f, dict): return f
     return {'type':'function', 'function':get_schema(f, pname='parameters')}
 
-# %% ../nbs/07_chat.ipynb #3e0afa31
-@dataclass
-class ToolResponse:
-    content: list[str,str]
-
 # %% ../nbs/07_chat.ipynb #bba6fd58
 def _mk_tool_result(res):
     "Unwrap `ToolResponse`, and format tool result message"

{python_fastllm-0.0.33 → python_fastllm-0.0.35}/fastllm/gemini.py

@@ -273,21 +273,18 @@ def get_hdrs(api_key=None):
 def cost(usage, m):
     raw = usage.raw
     prompt_tot = raw.get('promptTokenCount', 0)
-    tier = '_above_200k_tokens' if prompt_tot > 200_000 else ''
-    in_rate    = m.get(f'input_cost_per_token{tier}')       or m.input_cost_per_token
-    out_rate   = m.get(f'output_cost_per_token{tier}')      or m.output_cost_per_token
-    cache_rate = m.get(f'cache_read_input_token_cost{tier}')or m.get('cache_read_input_token_cost', 0)
-    audio_rate = m.get('input_cost_per_audio_token')  # None if not priced separately
-
+    tier = price_tier(m, prompt_tot)
+    in_rate    = tier_rate(m, 'input_cost_per_token', tier)
+    out_rate   = tier_rate(m, 'output_cost_per_token', tier)
+    cache_rate = tier_rate(m, 'cache_read_input_token_cost', tier)
+    # Models with no `input_cost_per_audio_token` (e.g. Gemini 3 Pro) bill audio at the standard input rate, so it stays in `in_txt`
+    audio_rate = m.get('input_cost_per_audio_token')
     cached = raw.get('cachedContentTokenCount', 0)
-    # Gemini 3 Pro supports bills audio at the standard input rate (no separate input_cost_per_audio_token key in the metadata)
     audio  = sum(d['tokenCount'] for d in raw.get('promptTokensDetails', []) if d.get('modality')=='AUDIO') if audio_rate else 0
     in_txt = prompt_tot - cached - audio
-
     thoughts = raw.get('thoughtsTokenCount', 0) or 0
     cands    = raw.get('candidatesTokenCount', 0) or 0
     reason_rate = m.get('output_cost_per_reasoning_token') or out_rate
-
     cost  = in_txt * in_rate + cands * out_rate
     cost += cached * cache_rate
     cost += audio  * (audio_rate or 0)

{python_fastllm-0.0.33 → python_fastllm-0.0.35}/fastllm/openai_responses.py

@@ -251,9 +251,10 @@ def get_hdrs(api_key=None):
 def cost(usage, m):
     raw = usage.raw
     cached = raw.get('input_tokens_details', {}).get('cached_tokens', 0)
-    in_txt = raw['input_tokens'] - cached
-    cost  = in_txt * m.input_cost_per_token + raw['output_tokens'] * m.output_cost_per_token
-    cost += cached * m.get('cache_read_input_token_cost', 0)
+    tier = price_tier(m, raw['input_tokens'])
+    cost  = (raw['input_tokens'] - cached) * tier_rate(m, 'input_cost_per_token', tier)
+    cost += raw['output_tokens'] * tier_rate(m, 'output_cost_per_token', tier)
+    cost += cached * tier_rate(m, 'cache_read_input_token_cost', tier)
     return cost
 
 # %% ../nbs/02_oai_responses.ipynb #07114b55

{python_fastllm-0.0.33 → python_fastllm-0.0.35}/fastllm/streaming.py

@@ -150,7 +150,7 @@ async def mk_acollect_stream(it, index_fn, model=None, api_name=None, vendor_nam
         deltas.append(d)
     part_accum.finalize()
     tcs = part_accum.tool_calls
-    if api_name: usg = api_registry.apis[api_name].finalize_usage(usg, part_accum.parts)
+    if api_name: usg = api_registry[api_name].finalize_usage(usg, part_accum.parts)
     if stop: fin = FinishReason.stop
     fin = FinishReason.tool_calls if fin==FinishReason.stop and any(~L(tcs).attrgot('server')) else fin # recheck tool calls post collation
     # tool calls and non-anthropic citations are yielded at the end

{python_fastllm-0.0.33 → python_fastllm-0.0.35}/fastllm/types.py

@@ -9,10 +9,11 @@ __all__ = ['FinishReason', 'api_registry', 'model_prices_url', 'haik45', 'sonn45
            'codex_pricing', 'sol', 'terra', 'luna', 'gpt56s', 'Usage', 'Completion', 'APIRegistry', 'mk_completion',
            'fn_schema', 'payload_kwargs', 'get_api_key', 'resize_b64', 'model_prices_meta', 'infer_api_name',
            'get_model_meta', 'register_model_info', 'get_model_info', 'get_model_pricing', 'approx_pricing',
-           'is_deepseek_peak_hour']
+           'is_deepseek_peak_hour', 'price_tier', 'tier_rate']
 
 # %% ../nbs/00_types.ipynb #b4d047fd
 import httpx, base64, io
+from importlib.metadata import entry_points
 from dataclasses import dataclass, field
 from datetime import datetime, timezone
 from fastcore.net import urljson
@@ -74,9 +75,18 @@ def _repr_markdown_(self: Completion):
 FinishReason = str_enum('finish_reason', 'stop', 'tool_calls', 'length', 'content_filter')
 
 # %% ../nbs/00_types.ipynb #fc681c52
-class APIRegistry:
-    def __init__(self): self.apis = {}
-    def register(self, name, finalize_usage=noop, **kwargs): self.apis[name] = SimpleNamespace(finalize_usage=finalize_usage, **kwargs)
+class APIRegistry(dict):
+    _eps = None
+    def register(self, name, finalize_usage=noop, **kwargs): self[name] = SimpleNamespace(finalize_usage=finalize_usage, **kwargs)
+    def _load(self, name):
+        "Import the plugin providing api `name` from the `fastllm.apis` entry-point group; its import registers the api."
+        if self._eps is None: self._eps = {o.name:o for o in entry_points(group='fastllm.apis')}
+        if name in self._eps: self._eps[name].load()
+        return dict.__contains__(self, name)
+    def __contains__(self, name): return dict.__contains__(self, name) or self._load(name)
+    def __missing__(self, name):
+        if self._load(name): return self[name]
+        raise KeyError(name)
 
 api_registry = APIRegistry()
 
@@ -85,7 +95,7 @@ api_registry = APIRegistry()
 def mk_completion(resp, model, api_name, vendor_name):
     "Normalize an api response into Completion."
     resp = obj2dict(resp)
-    api = api_registry.apis[api_name]
+    api = api_registry[api_name]
     tcs = api.norm_tool_calls(resp)
     parts = api.norm_parts(resp)
     usg = api.finalize_usage(api.norm_usage(resp), parts)
@@ -215,8 +225,9 @@ register_model_info('gemini-3.5-flash', vendor_name='gemini', base='gemini-3-fla
     input_cost_per_token=1.5e-6, output_cost_per_token=9e-6,
     output_cost_per_reasoning_token=9e-6, cache_read_input_token_cost=1.5e-7)
 
-for model in ('gpt-5.4', 'gpt-5.4-mini'):
-    register_model_info(model, vendor_name='openai', base=model, supports_web_search=True, mode=None, max_input_tokens=250000)
+register_model_info('gpt-5.4', vendor_name='openai', base='gpt-5.4', supports_web_search=True, mode=None)
+# Upstream metadata says 1,050,000 input tokens for gpt-5.4-mini, but OpenAI documents a 272k limit
+register_model_info('gpt-5.4-mini', vendor_name='openai', base='gpt-5.4-mini', supports_web_search=True, mode=None, max_input_tokens=272_000)
 
 for model in ('kimi-k2.5', 'kimi-k2.6'):
     register_model_info(model, vendor_name='moonshot', base=f'moonshot/{model}', base_vendor_name=None,
@@ -271,8 +282,14 @@ codex_pricing = dict(
     input_cost_per_token = 0.10/1_000_000, output_cost_per_token = 0.50/1_000_000,
     cache_creation_input_token_cost = 0.10/1_000_000, cache_read_input_token_cost = 0.10/1_000_000)
 
+def _rm_ctx_tiers(vendor_name, model):
+    "Drop upstream context-tier rates, which would otherwise override flat subscription pricing"
+    info = model_info_registry[vendor_name, model]
+    for k in [k for k in info if re.search(r'_above_\d+k_tokens', k)]: del info[k]
+
 for model in (codex54, codex54m, codex55):
     register_model_info(model, 'codex', base=model, base_vendor_name='chatgpt', supports_web_search=True, max_input_tokens=256000, **codex_pricing)
+    _rm_ctx_tiers('codex', model)
 
 register_model_info(codex53spark, 'codex', **codex_pricing,
     supports_vision=False, supports_image_input=False, supports_web_search=True, supports_reasoning=True, supports_function_calling=True,
@@ -285,8 +302,12 @@ for model in (haik45, sonn45, sonn46, sonn5, opus46, opus48, opus5, fable5):
 # %% ../nbs/00_types.ipynb #bb0c4c2a
 sol,terra,luna = gpt56s = 'gpt-5.6-sol gpt-5.6-terra gpt-5.6-luna'.split()
 for model in ['gpt-5.6']+gpt56s:
-    register_model_info(model, 'openai', base=model, base_vendor_name='openai', max_input_tokens=272_000)
-    register_model_info(model, 'codex', base=model, base_vendor_name='openai', max_input_tokens=272_000, **codex_pricing)
+    register_model_info(model, 'openai', base=model, base_vendor_name='openai')
+# Codex serves only the suffixed names; bare `gpt-5.6` is rejected with a ChatGPT account.
+# Its window is smaller than the API's: 371,331 input tokens is accepted and 371,981 is not, on all three.
+for model in gpt56s:
+    register_model_info(model, 'codex', base=model, base_vendor_name='openai', max_input_tokens=371_000, **codex_pricing)
+    _rm_ctx_tiers('codex', model)
 
 # %% ../nbs/00_types.ipynb #24cc47ec
 def get_model_pricing(mn, vendor_name, million=True):
@@ -310,11 +331,22 @@ def is_deepseek_peak_hour(dt=None):
     h = dt.hour + dt.minute/60
     return 1 <= h < 4 or 6 <= h < 10
 
+# %% ../nbs/00_types.ipynb #4c701619
+def price_tier(meta, n):
+    "Suffix of the pricing tier applying to `n` prompt tokens, or '' for base rates."
+    o = first(o for k in meta if (o:=re.match(r'input_cost_per_token(_above_(\d+)k_tokens)$', k)))
+    return o[1] if o and n>int(o[2])*1000 else ''
+
+# %% ../nbs/00_types.ipynb #956d2495
+def tier_rate(meta, key, tier):
+    "Rate for `key` at `tier`, falling back to the base key."
+    return meta.get(f'{key}{tier}') or meta[key]
+
 # %% ../nbs/00_types.ipynb #8bfca02d
 @patch(as_prop=True)
 def cost(self:Completion):
     meta = dict2obj(get_model_info(self.model, self.vendor_name))
-    api = api_registry.apis[self.api_name]
+    api = api_registry[self.api_name]
     if not hasattr(api, 'cost'): raise NotImplementedError(f"API: {self.api_name} doesn't have a registered `cost` function in ns")
     res = api.cost(self.usage, meta)
     return res*2 if self.vendor_name=='deepseek' and self.model.startswith('deepseek-v4') and is_deepseek_peak_hour() else res

{python_fastllm-0.0.33 → python_fastllm-0.0.35}/pyproject.toml

@@ -15,7 +15,7 @@ classifiers = [
     "Programming Language :: Python :: 3",
     "Programming Language :: Python :: 3 :: Only",
 ]
-dependencies = ['fastcore>=2.1.11', 'aidialog>=0.0.3', 'fastspec>=0.0.11', 'toolslm', 'pillow']
+dependencies = ['fastcore>=2.1.11', 'aidialog>=0.0.5', 'fastspec>=0.0.11', 'toolslm', 'pillow']
 
 [project.urls]
 Repository = "https://github.com/AnswerDotAI/fastllm"

{python_fastllm-0.0.33 → python_fastllm-0.0.35}/python_fastllm.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-fastllm
-Version: 0.0.33
+Version: 0.0.35
 Author-email: Kerem Turgutlu <keremturgutlu@gmail.com>
 License: Apache-2.0
 Project-URL: Repository, https://github.com/AnswerDotAI/fastllm
@@ -10,7 +10,7 @@ Classifier: Programming Language :: Python :: 3 :: Only
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 Requires-Dist: fastcore>=2.1.11
-Requires-Dist: aidialog>=0.0.3
+Requires-Dist: aidialog>=0.0.5
 Requires-Dist: fastspec>=0.0.11
 Requires-Dist: toolslm
 Requires-Dist: pillow

{python_fastllm-0.0.33 → python_fastllm-0.0.35}/python_fastllm.egg-info/requires.txt

@@ -1,5 +1,5 @@
 fastcore>=2.1.11
-aidialog>=0.0.3
+aidialog>=0.0.5
 fastspec>=0.0.11
 toolslm
 pillow

python_fastllm-0.0.33/fastllm/__init__.py

@@ -1 +0,0 @@
-__version__ = "0.0.33"