Default model routing (current flagships)¶
Decision record for which models Primr uses by default, why each tier stays version-pinned, and how the defaults move when providers ship new models.
Decision update (2026-11-14)¶
The project owner directed that no superseded model is selectable. Every
xAI tier uses grok-4.7, including writing and the xAI-only utility path.
Gemini Pro uses gemini-4-carbon, Google's strongest Gemini 4 model
(checked 2026-11-14), not Gemini 4 Argon and not the cheaper
gemini-2.5-pro auto-route target gemini-3.8-flash. Flash stays
gemini-3.8-flash. Lite stays gemini-3.5-flash-lite. OpenAI selections
use gpt-6-astra, gpt-6.1-sol, gpt-6.1-terra, or gpt-6-luna.
Anthropic selections use claude-fable-5-2, claude-opus-5-5,
claude-sonnet-5-5, or claude-haiku-5-5. Opt-in OpenAI images use
gpt-image-2.6-flare.
Superseded registry rows stay so historical quotes still resolve. New
requests for those ids are rejected, except the direct Gemini 2.5 rows
below. Embeddings were not swapped.
gemini-2.5-pro and gemini-2.5-flash are deprecated and shut down on
2026-12-01 (deprecations page checked 2026-11-16). Until that date an
explicit selection still runs and reports the deprecation. On that date
the selection is refused. The refusal names gemini-3.8-flash for Pro and
gemini-3.5-flash-lite for Flash, the provider routing targets, distinct
from the gemini-4-carbon default. Quotes before the shutdown keep the
stored 2.5 rates. Quotes on or after it use the successor rates. Direct
gemini-2.5-flash-lite is not deprecated and stays selectable. The
OpenRouter row for that id is not.
deep-research-preview-04-2026 has no named current replacement in the
catalogs checked on this date, so that row is unchanged.
xAI rates checked at https://docs.x.ai/developers/pricing on 2026-11-14:
grok-4.7 is $2 input, $0.50 cached input, and $6 output per 1M tokens
below 200k prompt tokens, and $4, $1, and $12 at or above 200k. The
analysis fallback chain no longer inserts grok-4.3 after grok-4.7.
Decision update (2026-10-28, previous flagship table)¶
The project owner directed that defaults track each provider's current flagship instead of waiting for a paid quality comparison per release. This is an owner decision, not a measured promotion: report quality on the new defaults has not been scored against the previous ones.
Sources checked on 2026-10-28: the xAI, Gemini, OpenAI, and Claude model catalogs, OpenAI pricing and deprecations, and the OpenRouter models API.
| Role | Previous default | Current default |
|---|---|---|
| Grok reasoning, every tier | grok-4.3 (MAX: grok-4.5) |
grok-4.7 |
| Grok writing, FAST and HYBRID | grok-4.20-non-reasoning |
unchanged (xAI's only non-reasoning model) |
| Gemini Flash (utility) | gemini-3-flash-preview |
gemini-3.8-flash (stable) |
| Gemini Pro (reasoning, premium) | gemini-3.1-pro-preview |
gemini-4-argon (stable) |
| Gemini writing | gemini-3.1-flash-lite |
gemini-3.5-flash-lite |
| OpenAI utility and writing | gpt-5.4-nano |
gpt-6-luna |
| OpenAI reasoning | gpt-5.6-terra |
gpt-6.1-terra |
| Anthropic utility and writing | claude-haiku-4-5 |
claude-haiku-5-5 |
| Anthropic reasoning | claude-sonnet-5 |
claude-sonnet-5-5 |
| OpenRouter utility / writing / reasoning | google/gemini-3.1-flash-lite / openai/gpt-4.1-mini / deepseek/deepseek-v3.2 |
google/gemini-3.5-flash-lite / openai/gpt-6-luna / deepseek/deepseek-v4.1-flash-1017 |
| LLM judge | grok-4.3 |
grok-4.7 |
Free estimator dry-runs (complete mode, no historical cache), before and after:
| Keys | Before | After |
|---|---|---|
| xAI + Gemini, HYBRID | $0.76 ($0.89 with AI Strategy) | $1.05 ($1.24) |
| xAI + Gemini, MAX | $8.53 ($9.93) | $8.53 ($9.93) |
| xAI only | $5.09 | $5.18 |
| Gemini only | $0.89 | $1.07 |
| OpenAI only | $1.05 | $1.24 |
| Anthropic only | $4.71 | $3.20 |
| Premium | $4.21 | $4.14 |
This is the documented price-floor change the promotion path below allows: the HYBRID default moves from about $0.76 to about $1.05. About $0.20 comes from Gemini 3.5 Flash-Lite writing output ($2.50 versus $1.50 per 1M) and about $0.11 from Grok 4.7 reasoning. Gemini 4 Argon rates rise to $4/$20 on 2026-12-01 and Gemini 3.8 Flash to $1.50/$7.50 on 2027-01-01; estimates now disclose such scheduled changes so an approval made before the date is not silently repriced. Grok 4.7 has a 500k context window, half of Grok 4.3's. This 2026-10-28 decision kept Grok 4.3 as the 1M-context analysis fallback. The 2026-11-14 decision removes that step.
Longer Strategic Overview section floors move this static estimator again. xAI plus Gemini HYBRID is about $1.08 ($1.26 with the AI Strategy). xAI-only is about $5.24 ($5.99 with the AI Strategy). MAX is about $8.65 ($10.05 with the AI Strategy). The OpenRouter default full run is about $0.68. The table above remains the 2026-10-28 flagship record.
Catalog recheck (2026-11-02): the xAI models page, the Gemini pricing page,
the OpenAI pricing page, and the Claude Sonnet 5.5 and Haiku 5.5 pages list
the same default IDs and rates. No default moves on that date.
The 2026-11-14 decision above moves the Pro default to gemini-4-carbon
and uses grok-4.7 for xAI writing.
grok-4.20-non-reasoning remains the documented alias of
grok-4.20-0309-non-reasoning.
Rules that still hold: tiers name exact versions, never provider aliases;
superseded models stay registered and priced for historical accounting and
explicit selection; retired models are marked unavailable with their
replacement; and tests/test_ai/test_routing.py fails if any key combination
routes a default role to a model outside the current flagship set. The
historical record below explains the earlier eval-gated policy.
Context¶
xAI shipped Grok 4.6 (API id grok-4.6) on August 12, 2026 as its current
flagship. Primr had already standardized on Grok 4.3 for hybrid reasoning
after the v1.24.0 cross-provider eval (sub-$1 default with Gemini writing), and
had pinned the explicit MAX tier to Grok 4.5.
The README demo of that period showed a dry-run for “Grok 4.3 hybrid”, matching the default CLI path at the time. The current demo shows Grok 4.7 hybrid.
API currency checkpoint (2026-08-13)¶
Official sources reviewed before any paid comparison:
- xAI model catalog and Grok 4.6 model card
- xAI pricing, cost tracking, and rate limits
- Responses migration, prompt cache affinity, and data retention
- Gemini latest models, Interactions API, and Deep Research
- OpenAI model catalog, GPT-5.6 guidance, and the Responses API
The resulting transport contract is explicit:
- Normal xAI text generation uses Responses, not deprecated Chat
Completions. Independent calls set
store=false; Primr does not need provider-side response retrieval for them. - Continuous sessions send full explicit history and a per-session
prompt_cache_key. They do not depend on server-side state, so local run artifacts remain the recovery authority. - xAI
cost_in_usd_ticksis recorded as exact billed cost. Token pricing is retained as the conservative fallback if any call omits exact cost. - Primr honors provider retry guidance with bounded exponential backoff,
jitter, and
Retry-After; billing exhaustion remains non-retryable. - Gemini Deep Research already uses background Interactions and File Search.
Direct Gemini generation may migrate from supported legacy
generateContentseparately, with behavior tests rather than as an incidental change to the paid quality baseline. - OpenAI text generation uses Responses with
store=false. GPT-5.6 Sol/Terra/Luna and thegpt-5.6Sol alias are registered at current pricing as evaluation candidates; existing OpenAI production routing remains pinned until a representative comparison supports promotion.
Measured dry-run (local estimator, Aug 2026)¶
Keys = XAI_API_KEY + GEMINI_API_KEY, mode = complete/full, no historical cache:
| Tier | Model shape | Approx total |
|---|---|---|
fast / hybrid (default) |
4.3 reasoning + Gemini Flash-Lite writing + Gemini Flash utility | ~$0.76 |
max |
4.5 for every Grok-priced stage | ~$8.53 |
xAI-only (no Gemini) hybrid is ~$5.09; max is still ~$8.53.
Published list prices (docs.x.ai):
| Model | In / out (per 1M) | Context | Cached in (base) |
|---|---|---|---|
grok-4.3 |
$1.25 / $2.50 (≥200k: $2.50 / $5.00) | 1M | $0.20 |
grok-4.20 variants |
$1.25 / $2.50 (>=200k: $2.50 / $5.00) | 1M | $0.20 (>=200k: $0.40) |
grok-4.5 |
$2.00 / $6.00 (>=200k: $4.00 / $12.00) | 500k | $0.30 (>=200k: $0.60) |
grok-4.6 |
$2.00 / $6.00 (>=200k: $4.00 / $12.00) | 500k | $0.50 (>=200k: $1.00) |
4.5 is roughly 1.6× input / 2.4× output versus 4.3 at the standard tier, with a smaller context window. MAX is expensive mainly because it routes writing through Grok as well, not only reasoning.
Decision (2026-08, superseded on 2026-10-28)¶
- Default hybrid/fast reasoning stays
grok-4.3. The product promise is a cost-aware brief with an honest dry-run, not “always the newest flagship.” --grok-tier maxremains pinned togrok-4.5. MAX is an explicit, versioned production route, not an alias that changes when a provider ships a model. Re-estimate before spend.grok-4.6is registered as available but is not routed automatically. This keeps dry-run and output behavior stable while making the current model available to the evaluation harness.- Do not promote 4.5 or 4.6 to hybrid or MAX without an eval gate.
MODEL_ONBOARDING.mdstep 5 still applies: utility-per-dollar ≥ current default, hallucination/drift not worse. - Demo / README keep “Grok 4.3 hybrid.” Showing a flagship on the default demo would misrepresent cost and routing.
Why this is not “stale”¶
xAI recommends 4.6 for general code and knowledge work, but provider recency is not evidence of Primr report quality. Primr's heavy stages are long-form research writing and cross-validation with large prompts. The 4.6 cached-input rate is higher than 4.5, and both become substantially more expensive at the 200k boundary. Silent model aliases would make estimates and reproducibility less trustworthy.
Promotion path (when quality justifies it)¶
- Use the registered
grok46-flashliteprofile: 4.6 reasoning plus Gemini Flash-Lite writing, not max-everywhere. - Run the standing company corpus: hybrid (4.3) vs candidate (4.6 reasoning)
with the same keys and mode; score with
primr eval. - Pre-commit criteria:
- End-to-end dry-run still within the product’s “default under ~$1” envelope or a deliberate, documented price floor change.
- Trust / calibration / leakage metrics ≥ current hybrid.
- Only then flip
GROK_MODEL/ hybridget_grok_modelsand refresh the demo.
Until that passes, 4.6 remains a registered candidate. Grok 4.5 remains the MAX and analysis-fallback pin, not a claim that it is the newest model.
Explicitly not¶
- Silent default swap because a screenshot said “4.3.”
- Treating "newest xAI model" as the hybrid or MAX route without cost evidence.
- Using MAX-everywhere cost (~$8+) as the public default estimate.
- Region or account availability assumptions without checking the operator's visible model list.
Validation cost¶
- Free: registry unit tests, estimator tier labels, architecture ceilings.
- Paid: any hybrid-versus-4.6 reasoning eval on real companies. It shares the current $25 aggregate evaluation ceiling unless the operator explicitly authorizes a later, separate campaign. Never run it merely because the model was registered.