Model catalogs, prices, free-API terms and data policies were rechecked on 2026-10-29. I would still recheck both the public docs and account limits before integrating.
When choosing a hosted model, pick the tier first and the vendor second. I read the catalogs of five model vendors (OpenAI, Anthropic, Google, xAI, DeepSeek), and each line now sorts into three tiers: flagships handle the hard cases, mid-tier models carry routine volume, and small models run repetitive work cheaply. Image, video, and realtime voice are also moving toward dedicated endpoints.
The tiers themselves are clear; the line between them got harder to draw in late September. OpenAI says GPT-6.1 Sol comes close to its flagship Astra while listing at a fifth of Astra's price: 2 USD in and 10 USD out per million tokens, against Astra's 10 and 50. Anthropic's Opus 5.5 lists for less than Opus 5 did, and Sonnet 5.5 keeps Sonnet 5's price. Taking the vendors' claims at face value, a mid-tier price now buys close to flagship work, so the tier decision depends less on list price than on latency, effort settings, and how often a task still fails.
So I no longer start with “which company is best?” I start with three questions: does this job deserve a flagship, can a mid-tier model handle it, and is the extra control of open weights worth operating the stack myself?
GPT-6 Astra / GPT-6.1 Sol / GPT-6.1 Terra / GPT-6 Luna
The catalog's flagship lineup now lists four models, alongside specialized ones. Astra remains the flagship; GPT-6.1 Sol, released at DevDay on September 29, is described as “near-Astra performance for complex work at a lower cost”; GPT-6.1 Terra is “a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing”; Luna covers focused, high-volume work. All four list 1.05M context and 128K output, but capacity alone says little about reliability on long documents. The GPT-5.6 Sol / Terra / Luna line no longer appears among current models.
Claude Fable 5.1 / Opus 5.5 / Sonnet 5.5 / Haiku 5.5
Opus 5.5 (September 22) and Sonnet 5.5 (September 28) succeed Opus 5 and Sonnet 5, which are now legacy. Anthropic suggests starting most workloads on Opus 5.5 and moving to Fable 5.1 when Opus 5.5 at higher effort still falls short. Haiku 5.5, released October 7, succeeds Haiku 4.5, from USD 0.10 input and 0.50 output per million tokens; Haiku 4.5 is now legacy.
Gemini 4 Argon / Gemini 3.8 Flash (stable) / Gemini 3.1 Pro (preview), plus Live, TTS, image, and video models
The split by modality is especially clear. For realtime speech or native media, I would test the dedicated endpoint instead of making a text flagship do everything. Gemini 4 Argon, announced September 30, first went only to cyber defenders in the Fairwind Program and trusted testers; since October 13 it is generally available in the Gemini API as gemini-4-argon, though not on the Free tier. On October 8, gemini-3.7-flash and gemini-3.5-flash were deprecated; requests to them are routed automatically to 3.8 Flash and 3.6 Flash. Since September 18, Gemini 2.5 models remain available only to users who had already been using them.
Grok 4.7, released in September, is positioned for coding, agentic tasks, and knowledge work, with a 500K context window; 4.6 remains available. Current events still require search tools; they are not baked into the model.
deepseek-flash serves the snapshot DeepSeek-V4.1-Flash-1017, including vision; deepseek-v4-pro serves V4 Pro-0813 without vision. Both support thinking/non-thinking modes, Responses API, and OpenAI-/Anthropic-format access.
The legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp IDs remain accepted but are now served by V4.1-Flash-1017 at Flash prices; they no longer select the retired models. DeepSeek's server-side Search and Fetch tools cost 0.75 USD per 1,000 successful requests each. Calls made through DeepSeek Harness, Claude Code or Codex are free for now; per-request billing for them starts at 00:00 Beijing time on November 1, 2026.
In practice I would start in the middle tier. GPT-6.1 Sol, Sonnet 5.5, Gemini 3.8 Flash, Grok 4.7, and DeepSeek V4.1 Flash / V4 Pro can all enter the first test; GPT-6.1 Terra, at half Sol's price, joins that round when the work is screenshots, computer use, or writing. I would test Astra, Fable 5.1, Opus 5.5, or Gemini 4 Argon only after repeated task failures justify the cost. Putting Opus 5.5 there departs from Anthropic's advice: Anthropic tells people who are unsure to start most workloads on Opus 5.5, while I sort by price. Opus 5.5 lists at 4 / 20 USD per million tokens, twice Sonnet 5.5's 2 / 10, so the first round runs on Sonnet 5.5 and only the tasks it fails go to Opus 5.5. In the other direction, extraction, classification, and format conversion that already work reliably should keep moving toward Luna or another cheaper tier; on Anthropic’s side, that tier is Haiku 5.5. When the output is only a choice among a few options, a decision model may be cheaper still: OpenAI's Decisions API runs on gpt-6-luna at 0.10 USD per million uncached input tokens, and Jev lists 0.042 USD; neither charges for output.
That is cheaper than throwing a flagship at every problem and less frustrating than optimizing for the lowest token price from day one.
Free APIs: keep a shortlist, not a permanent leaderboard
For a personal prototype, I now keep three main routes: Gemini for long context and multimodal work, Groq for low latency, and OpenRouter for trying several free models behind one interface. All can start at zero payment, but they differ sharply in units, reset windows, and data terms.
Service
Current zero-payment entry
How I would use it
Boundary to check first
Gemini Developer API
A Free tier for models such as Gemini 3.8 Flash; AI Studio is free in supported regions; Gemini 3.1 Pro has no free API tier
Long documents, multimodal work, and research prototypes
Quotas vary by model and account. Unpaid content may be used to improve Google products and read by human reviewers (see data terms below), so sensitive material stays out
Groq
Free Plan with organization- and model-level RPM, RPD, TPM, and TPD limits
Low-latency chat, voice, and tool workers
The public page is a summary; the account Limits page is the exact current authority
OpenRouter :free
Multi-provider free models behind one compatible API
Free mode enables API access without a credit card
Mistral model prototypes
The public promise is limited usage and rate limits, not a stable numeric allowance
Cloudflare Workers AI
10,000 Neurons free per day, resetting at 00:00 UTC
Edge tasks already living on Cloudflare
Neurons are not tokens. Eight models need the Workers Paid plan or prepaid credits: Kimi K2.6 and K2.7-code, GLM-5.2, GLM-5.3 and GLM-5.3-flash, and the DeepSeek V4 Flash, V4.1 Flash and V4 Pro builds
Zhipu GLM-4.7-Flash
Still documented under the free-model path with an OpenAI-style chat-completions endpoint
A Chinese and agentic-coding alternative
The page gives 200K context and 128K maximum output, but no longer states free use or any quota; confirm the price in the console before relying on it
The most useful update is not another provider joining the list, but three leaving it. Cerebras now offers a 5 USD, 30-day Free Trial after a verified payment method, not an automatically renewing free tier. GitHub Models was fully retired on 2026-07-30.Hugging Face Inference Providers no longer gives Free accounts monthly credits: they must buy credits, while PRO accounts get 2 USD a month to spend across Hugging Face compute.
I therefore would not build around one free endpoint. A personal project can start with OpenRouter for interface consistency, then keep direct Gemini and Groq integrations for clearly assigned workloads. Cloudflare or GLM only joins when its ecosystem or language advantage matters. Every route still needs 429 handling, Retry-After, a daily budget, and failover; none should receive private code or customer data by default.
This section is about API inference allowance, not usage bundled into terminal agents. That separate accounting problem lives in Free Allowances for AI Coding CLIs.
Yes: under the terms, unpaid content may be used to improve Google products and read by human reviewers
Users in the EEA, Switzerland and the UK get the paid-service data terms even on unpaid quota
Gemini API, paid tier
No
Prompts and responses are logged for a limited period for abuse detection. Since October 26, new billing-enabled projects also log Generate Content requests by default; logs are kept up to 55 days (7, 14 or 28 can be chosen), are not used for training unless you share a dataset, and can be turned off in AI Studio
xAI
No, not without explicit permission (security FAQ)
Requests and responses stored for 30 days for abuse auditing
When data locality or stable volume justifies running a service, move to the local AI stack. Hardware fit, coding candidates, runtimes, and evaluation each have a separate owner there; this hosted radar does not duplicate their snapshots.
These are catalog rates checked on 2026-10-29, in USD per million input/output tokens for standard text usage, not total task costs or account entitlements. Caching, batch jobs, reasoning, long context, and contracts can change the actual bill substantially.
Vendor
Current tiers (input / output)
OpenAI
Astra 10 / 50; GPT-6.1 Sol 2 / 10; GPT-6.1 Terra 1 / 5; Luna 0.10 / 0.50
Gemini 3.8 Flash: 0.75 / 3.75 through December 31, 2026, then 1.50 / 7.50 from January 1, 2027; Gemini 4 Argon: 2 / 10 through November 30, then 4 / 20 from December 1
xAI
Grok 4.7 and 4.6: 2 / 6 below 200K prompt tokens; 4 / 12 at or above 200K
I would not switch models from this table alone. A call that costs half as much can still be more expensive if it needs retries, produces repairs, or keeps failing at tool use. The useful number is the cost of finishing one real task, not the sticker price of one token.
An agent loop resends the same prefix on every turn: instructions, tool definitions, the files it has already read. Once that prefix is cached, providers bill it at a fraction of the input price, and work that can wait for a batch job costs less again. For an agent, these two discounts often move the bill more than a change of tier.
Cache reads cost 5% on Opus 5.5 and Sonnet 5.5, 2.5% on Fable 5.1, 10% elsewhere; writing the cache costs 1.25× base input for 5 minutes or 2× for 1 hour
Batch 50% off input and output; it stacks with caching
deepseek-flash cache hit 0.003 against 0.15 off-peak (2%)
Off-peak costs half of peak; peak is 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, except Chinese public holidays
A worked example on Opus 5.5, ignoring output: a 20-turn session resends the same 50,000-token prefix each turn. Uncached, that is 1,000,000 input tokens at 4 USD per million, so 4.00 USD. With a 5-minute cache, the first turn writes 50,000 tokens at 5 USD per million (0.25 USD) and the other 19 turns read 950,000 tokens at 0.20 USD per million (0.19 USD): 0.44 USD in total. Moving the uncached session down to Sonnet 5.5 only halves it, to 2.00 USD. A cache hit needs the same prefix, so the stable part of the prompt goes first and the changing part last.
The vendor material is enough to shrink the shortlist, but it is not a shared benchmark. I am not going to crown an overall winner from product pages. The useful next step is to replay the same work on my own repositories and documents, with failure criteria that matter to me. Only then is a default model actually personal.