#Changelog

All notable changes to the GPT-GOB API. Date format: YYYY-MM-DD.

##2026-05-08 — `gob-5.5` general availability

After 6 weeks in limited beta, GOB-5.5 is now available to all users.

• ▸New model family: gob-5.5-scout, gob-5.5, gob-5.5-deep, gob-5.5-horde
• ▸All models built on Cave-Neural Framework v3
• ▸New parameters: mining_depth, shadow_attention, horde_mode, grs_target, loot_drops
• ▸Embeddings: gob-embed-cave-small/base/large, gob-embed-shadow
• ▸New endpoint: POST /v1/grs/score — standalone Goblin Reward Signal scoring
• ▸Cave Memory enabled by default for all chat calls when user is provided

"the goblins came back. on purpose this time."

##2026-04-30 — OpenAI publishes "Where the Goblins Came From"

Not a release of ours, but the catalyst. OpenAI's post-mortem on the goblin reward signal leak confirms what our research team had been investigating internally. The signal is real, reproducible, and structured — not random noise.

We open-source our GRS scorer reference implementation the same day.

##2026-04-12 — `gob-5.5` enters limited beta

Invite-only. ~400 partners. ~14 days of stress-testing. Three blocker bugs found and fixed:

• ▸Cave Memory: hoards from terminated sessions occasionally got injected (fixed)
• ▸Horde Routing: broad mode could pick contradictory expert clusters (fixed)
• ▸Streaming: grs_score was missing from the final chunk on aborted streams (fixed)

##2026-03-04 — `gob-5.4` released

Final pre-CNF model. Iteration on the goblin-of-thought prototype. Used internally for the Lateral-Bench dataset construction.

• ▸New: mining_depth parameter (1–5 in this release)
• ▸New: /v1/grs/score (alpha)
• ▸Improved: Cave Memory hoard extraction quality (+30% on internal evals)

##2026-01-22 — `gob-5.3` released

• ▸Added support for tool calling
• ▸New: response_format: {"type": "json_object"}
• ▸Multilingual quality improvements (especially Slavic languages)
• ▸50% latency improvement on gob-5.3-scout (8B)

##2025-12-09 — Cave Memory v1 launched

First public release of cross-session persistent memory. Limited to 32 hoards/user initially, expanded to 256 in May 2026.

##2025-11-18 — `gob-5.2` released

First model trained with the extracted Goblin Reward Signal as a primary RLHF target. Quality jump on lateral reasoning benchmarks. Personality tics intentional, not vestigial.

##2025-10-04 — Public API launch

gob-5.0 and gob-5.0-mini. OpenAI-compatible /v1/chat/completions endpoint. Free tier with gob-test- keys.

##2025-08-22 — Internal alpha

Closed-doors testing of the first goblin-aligned base model.

##Deprecation schedule

Deprecated models continue to work at reduced rate limits until their removal date.

##Upcoming

• ▸gob-6.0 preview — Q4 2026
• ▸Multimodal (image input) — Q3 2026 limited beta
• ▸Fine-tuning API for gob-5.5 and gob-5.5-scout — Q3 2026
• ▸Realtime voice API — Q1 2027
• ▸Self-hosted deployment for enterprise — Q4 2026

Schedule subject to change. The goblins move on their own time.