python-fastllm 0.0.34__tar.gz → 0.0.36__tar.gz

{python_fastllm-0.0.34 → python_fastllm-0.0.36}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-fastllm
-Version: 0.0.34
+Version: 0.0.36
 Author-email: Kerem Turgutlu <keremturgutlu@gmail.com>
 License: Apache-2.0
 Project-URL: Repository, https://github.com/AnswerDotAI/fastllm

python_fastllm-0.0.36/fastllm/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.0.36"

{python_fastllm-0.0.34 → python_fastllm-0.0.36}/fastllm/_modidx.py

@@ -95,7 +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'),
@@ -258,6 +257,7 @@ d = { 'settings': { 'branch': 'main',
                                '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'),
@@ -269,5 +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-0.0.34 → python_fastllm-0.0.36}/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
@@ -18,7 +18,7 @@ from dataclasses import dataclass
 
 from .types import *
 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, fmt2hist,
+    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 *
 
@@ -141,11 +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-0.0.34 → python_fastllm-0.0.36}/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-0.0.34 → python_fastllm-0.0.36}/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-0.0.34 → python_fastllm-0.0.36}/fastllm/types.py

@@ -9,7 +9,7 @@ __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
@@ -225,10 +225,11 @@ 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'):
+for model in ('kimi-k2.5', 'kimi-k2.6' , 'kimi-k3'):
     register_model_info(model, vendor_name='moonshot', base=f'moonshot/{model}', base_vendor_name=None,
         supports_reasoning=True, supports_vision=True, supports_assistant_prefill=True)
 
@@ -241,9 +242,11 @@ register_model_info('accounts/fireworks/models/kimi-k2p5', vendor_name='firework
 register_model_info('accounts/fireworks/models/kimi-k2p6', vendor_name='fireworks_ai', base='accounts/fireworks/models/kimi-k2p5',
     supports_reasoning=True, supports_vision=True,
     input_cost_per_token=0.95e-6, cache_read_input_token_cost=0.16e-6, output_cost_per_token=4.0e-6)
-
 register_model_info('kimi-k2.7-code', vendor_name='moonshot', base='kimi-k2.6',
     input_cost_per_token=0.95e-6, cache_read_input_token_cost=0.19e-6, output_cost_per_token=4.0e-6)
+register_model_info('kimi-k3', vendor_name='moonshot', base='kimi-k2.6',
+    supports_reasoning=True, supports_vision=True, max_input_tokens=1_000_000,
+    input_cost_per_token=3e-6, cache_read_input_token_cost=0.30e-6, output_cost_per_token=15.0e-6)
 
 # %% ../nbs/00_types.ipynb #948d55d0
 deepseek_v4_common = dict(**modern_llm, supports_assistant_prefill=True,
@@ -281,8 +284,14 @@ 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,
@@ -295,8 +304,12 @@ 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):
@@ -320,6 +333,17 @@ 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):

{python_fastllm-0.0.34 → python_fastllm-0.0.36}/python_fastllm.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-fastllm
-Version: 0.0.34
+Version: 0.0.36
 Author-email: Kerem Turgutlu <keremturgutlu@gmail.com>
 License: Apache-2.0
 Project-URL: Repository, https://github.com/AnswerDotAI/fastllm

python_fastllm-0.0.34/fastllm/__init__.py

@@ -1 +0,0 @@
-__version__ = "0.0.34"